WO2021082231A1 - 目标检测方法及装置、电子设备和存储介质 - Google Patents

目标检测方法及装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2021082231A1
WO2021082231A1 PCT/CN2019/127181 CN2019127181W WO2021082231A1 WO 2021082231 A1 WO2021082231 A1 WO 2021082231A1 CN 2019127181 W CN2019127181 W CN 2019127181W WO 2021082231 A1 WO2021082231 A1 WO 2021082231A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
processing
image
target detection
level
Prior art date
Application number
PCT/CN2019/127181
Other languages
English (en)
French (fr)
Inventor
颜鲲
杨昆霖
侯军
伊帅
Original Assignee
北京市商汤科技开发有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京市商汤科技开发有限公司 filed Critical 北京市商汤科技开发有限公司
Priority to JP2022519510A priority Critical patent/JP2022549728A/ja
Priority to KR1020227009421A priority patent/KR20220050960A/ko
Publication of WO2021082231A1 publication Critical patent/WO2021082231A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10004—Still image; Photographic image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20112—Image segmentation details
    • G06T2207/20164—Salient point detection; Corner detection

Definitions

  • the present disclosure relates to the field of computer vision technology, and in particular to a target detection method and device, electronic equipment and storage medium.
  • Computer vision is a technology that uses computers and related equipment to simulate biological vision.
  • the three-dimensional information of the corresponding scene can be obtained by processing the collected images or videos.
  • the collected images or videos can be used for target detection to determine the target object category and position in the image.
  • the target detection technology can use neural networks to directly determine the target object's category and location detection frame.
  • the present disclosure proposes a technical solution for target detection.
  • a target detection method including: acquiring a detection image to be processed; determining, based on the image characteristics of the detection image, the size feature and corner point feature corresponding to the target detection object; The feature and the corner point feature are used to extract the object feature corresponding to the target detection object from the image feature; the category of the target detection object is determined based on the object feature.
  • the determining the size feature and corner point feature corresponding to the target detection object according to the image feature of the detection image includes: performing at least one level of convolution processing on the detection image to obtain the The image feature of the detection image; corner pooling processing is performed on the image feature of the detection image to obtain the size feature and the corner feature corresponding to the target detection object.
  • the convolution processing includes up-sampling processing and down-sampling processing; the performing at least one level of convolution processing on the detection image to obtain the image characteristics of the detection image includes: The detection image is subjected to at least one level of down-sampling processing to obtain a first feature map after at least one level of down-sampling processing; based on the first feature map after at least one level of up-sampling processing, at least one level of down-sampling processing is obtained The second feature map; based on the first feature map after the at least one level of down-sampling processing and the second feature map after the at least one level of up-sampling processing, the image features of the detection image are obtained.
  • each level of the down-sampling process outputs a first feature image
  • each level of the up-sampling process outputs a second feature image
  • Sampling the first feature map after sampling processing to obtain the second feature map after at least one level of downsampling processing includes: for the first level of upsampling processing in the at least one level of upsampling processing, subtracting the at least one level of downsampling processing The first feature map after the last downsampling process in the sampling process is used as the input of the first upsampling process; the second feature map output after the first upsampling process is obtained; for the at least one In the Nth-stage upsampling process, the Nth-stage upsampling processing is performed by combining the second feature map output after the upper-stage up-sampling processing of the N-th stage The first feature map of the second feature map is used as the input of the Nth upsampling process; the second feature map output by the Nth upsampling process is obtained, where N is a
  • the second feature map output after the upper-stage up-sampling processing of the Nth-stage up-sampling processing and the second feature map outputted after the upper-stage up-sampling processing of the Nth-stage upsampling processing are matched
  • the first feature map of the feature map, as the input of the Nth level of upsampling processing includes: the second feature map output after the upper level of upsampling processing of the Nth level of upsampling processing is matched with the Feature fusion is performed on the first feature map of the second feature map output after the N-th level upsampling processing to obtain the input of the N-th level upsampling processing.
  • the performing corner pooling processing on the image features of the detected image to obtain the size feature and corner feature corresponding to the target detection object includes: performing the image feature of the detected image Corner pooling processing to obtain the processing result; using the first branch network to perform convolution processing on the processing result to obtain the size feature corresponding to the target detection object; using the second branch network to perform convolution processing on the processing result to obtain The corner point feature corresponding to the target detection object; wherein, the number of channels of the first branch network and the second branch network are different.
  • the extracting the object feature corresponding to the target detection object from the image feature based on the size feature and the corner feature includes: according to the size feature and the The corner point feature determines a feature area that has a mapping relationship with the image area of the target detection object in the detection image; and extracts the object feature corresponding to the target detection object in the feature area of the image feature.
  • the corner feature corresponding to the target detection object includes at least a first corner feature and a second corner feature corresponding to the target object;
  • the size feature corresponding to the target detection object includes: The length feature and width feature of the target detection object corresponding to the first corner feature, and the length feature and width feature of the target detection object corresponding to the second corner feature.
  • the method further includes: determining the target based on the length feature and width feature corresponding to the first corner point feature and the length feature and width feature corresponding to the second corner point feature The detection frame of the detection object in the detection image; determining the intersection ratio between any two overlapping detection frames; in the case that the intersection ratio is greater than a preset threshold, the any two overlapping The detection frame is combined into one detection frame.
  • the determining the category of the target detection object based on the object feature includes: performing at least one level of convolution processing on the object feature to obtain the target detection object The probability of belonging to at least one preset category; according to the probability that the target detection object belongs to at least one preset category, the category of the target detection object is determined from the preset category.
  • a target detection device including: an acquisition module for acquiring a detection image to be processed; a determination module for determining the size of the target detection object according to the image characteristics of the detection image Features and corner features; an extraction module for extracting the object features corresponding to the target detection object from the image features based on the size features and the corner features; a classification module for extracting object features based on the object features Determine the category of the target detection object.
  • the determining module is specifically configured to perform at least one level of convolution processing on the detection image to obtain the image characteristics of the detection image; and perform corner points on the image characteristics of the detection image. Pooling process to obtain the size feature and corner feature corresponding to the target detection object.
  • the convolution processing includes up-sampling processing and down-sampling processing; the determining module is specifically configured to perform at least one level of down-sampling processing on the detection image to obtain at least one level of down-sampling processing.
  • the first feature map and the second feature map after the at least one level upsampling process are used to obtain the image feature of the detection image.
  • each level of the down-sampling process outputs a first feature image
  • each level of the up-sampling process outputs a second feature image
  • the determining module is specifically configured to: Regarding the first-stage up-sampling processing in the at least one-stage up-sampling processing, the first feature map after the last-stage down-sampling processing in the at least one-stage down-sampling processing is used as the first-stage up-sampling processing
  • the second feature map output after the first-level upsampling processing is obtained; for the N-th up-sampling processing in the at least one-level up-sampling processing, the previous one of the N-th upsampling processing is
  • the second feature map output after the stage upsampling processing and the first feature map matching the second feature map output after the Nth stage upsampling processing are used as the input of the Nth stage upsampling processing; to obtain the The second feature map output by the N-th upsampling process, where N is a positive integer greater than
  • it is specifically used to compare the second feature map output after the upper-level up-sampling processing of the N-th level up-sampling processing with the second feature map output after the upper-level up-sampling processing of the N-th stage up-sampling processing. Perform feature fusion on the first feature map of the second feature map to obtain the input of the Nth level upsampling processing.
  • the determining module is specifically configured to perform corner pooling processing on the image features of the detected image to obtain a processing result; use the first branch network to convolve the processing result Processing to obtain the size feature corresponding to the target detection object; use the second branch network to perform convolution processing on the processing result to obtain the corner feature corresponding to the target detection object; wherein the number of channels of the first branch network and the second branch network different.
  • the extraction module is specifically configured to determine, based on the size feature and the corner point feature, that there is a mapping relationship with the target detection object in the image area of the detection image Feature area; extract the object feature corresponding to the target detection object in the feature area of the image feature.
  • the corner feature corresponding to the target detection object includes at least a first corner feature and a second corner feature corresponding to the target object;
  • the size feature corresponding to the target detection object includes: The length feature and width feature of the target detection object corresponding to the first corner feature, and the length feature and width feature of the target detection object corresponding to the second corner feature.
  • the device further includes: a merging module, configured to be based on the length feature and width feature corresponding to the first corner point feature and the length feature and width feature corresponding to the second corner point feature , Determine the detection frame of the target detection object in the detection image; determine the intersection ratio between any two overlapping detection frames; when the intersection ratio is greater than a preset threshold, combine the Any two overlapping detection frames are merged into one detection frame.
  • a merging module configured to be based on the length feature and width feature corresponding to the first corner point feature and the length feature and width feature corresponding to the second corner point feature , Determine the detection frame of the target detection object in the detection image; determine the intersection ratio between any two overlapping detection frames; when the intersection ratio is greater than a preset threshold, combine the Any two overlapping detection frames are merged into one detection frame.
  • the classification module is specifically configured to perform at least one level of convolution processing on the object feature to obtain the probability that the target detection object belongs to at least one preset category; according to the target The detection object belongs to the probability of at least one preset category, and the category of the target detection object is determined from the preset category.
  • an electronic device including: a processor; a memory for storing executable instructions of the processor; wherein the processor is configured to execute the above-mentioned target detection method.
  • a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the foregoing target detection method is realized.
  • a computer program wherein the computer program includes computer-readable code, and when the computer-readable code runs in an electronic device, a processor in the electronic device executes To achieve the above-mentioned target detection method.
  • Fig. 1 shows a flowchart of a target detection method according to an embodiment of the present disclosure.
  • Fig. 2 shows a block diagram of a detection frame of a target detection object according to an embodiment of the present disclosure.
  • Fig. 3 shows a block diagram of using a neural network to obtain a target detection object category according to an embodiment of the present disclosure.
  • Fig. 4 shows a block diagram of a target detection device according to an embodiment of the present disclosure.
  • Fig. 5 shows a block diagram of an example of an electronic device according to an embodiment of the present disclosure.
  • the target detection solution provided by the embodiments of the present disclosure can obtain the detection image to be processed, and then determine the size feature and corner point feature corresponding to the target detection object according to the image feature of the detection image, and then based on the size feature and corner point feature, it can be obtained from From the image feature, the object feature corresponding to the target detection object is extracted, and then based on the object feature corresponding to the target detection object, the category of the target detection object is determined, so that the detection result of the target detection can be obtained.
  • the object characteristics corresponding to the target detection object can be determined first, and then the target detection objects can be classified according to the corresponding object characteristics, so that the object characteristics and target objects of the target detection object can be determined non-parallel in the target detection
  • the category makes the detection result obtained more accurate and can improve the accuracy of the detection result.
  • target detection usually requires intensive collection of anchor points forming a preliminary detection frame, but there will be many invalid anchor points in a large number of anchor points, which will consume processing time and storage space.
  • the detection frame and category of the target detection object are determined in parallel, so that the information of the detection frame cannot be considered when determining the category of the target detection object, resulting in insufficient detection results.
  • the target detection solution provided by the embodiments of the present disclosure can first determine the object characteristics of the target detection object by determining the corner points and the size of the target detection object, thereby reducing the waste of time and storage space caused by collecting a large number of anchor points, and
  • the object features obtained through the corner points and the size can reduce the difficulty of distinguishing different target detection objects through the center point determined by the anchor point when two target detection objects are found to be coincident, so that the embodiments of the present disclosure provide
  • the target detection scheme can distinguish different target detection objects by corner features, which can save time and storage space for acquiring anchor points, improve the efficiency of obtaining detection results, and obtain detection results with higher accuracy.
  • the information processing solution provided by the embodiments of the present disclosure can be applied to any scene that requires target detection.
  • it can be applied to the scene of performing target detection on the collected video, and the trajectory of the target detection object in the video can be obtained according to the detection result.
  • it can be applied to security scenarios, and suspects can be identified and tracked based on the detection results.
  • the target detection solution provided by the present disclosure will be described below through embodiments.
  • Fig. 1 shows a flowchart of a target detection method according to an embodiment of the present disclosure.
  • the target detection method can be executed by a terminal device, server or other target detection device, where the terminal device can be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital processing ( Personal Digital Assistant, PDA), handheld devices, computing devices, vehicle-mounted devices, wearable devices, etc.
  • the target detection method may be implemented by a processor invoking computer-readable instructions stored in the memory.
  • the target detection method in the embodiment of the present disclosure will be described below by taking the target detection device as the execution subject as an example.
  • the target detection method includes the following steps:
  • Step S11 Obtain a detection image to be processed.
  • the target detection device may have an image acquisition function, and may shoot the current scene to obtain the detected image to be processed.
  • the target detection apparatus may obtain the detection image to be processed by other equipment.
  • the detection image may be a separately collected image, or may be an image frame in the video stream.
  • the acquired detection image to be processed may be a preprocessed detection image, for example, may undergo preprocessing such as image scaling, image enhancement, and image filtering.
  • preprocessing such as image scaling, image enhancement, and image filtering.
  • the length and width of the detected image can be adjusted to an appropriate size to obtain a pre-processed detected image.
  • Step S12 Determine the size feature and corner point feature corresponding to the target detection object according to the image feature of the detection image.
  • a neural network can be used to extract the image features of the detection image. From the image features of the detection image, the size feature corresponding to the target detection object can be determined, and the corner feature corresponding to the target detection object can be determined.
  • the target detection object may be an object that needs to be detected in the detection image, for example, an image of a pedestrian, a vehicle, a building, a sign, etc.
  • the size feature of the target detection object can represent the size feature of the image area where the target detection object is located.
  • the size feature corresponding to the target detection object may be the length feature and/or the width feature corresponding to the quadrilateral.
  • the corner point feature can characterize the position information of the diagonal point of the image area where the target detection object is located.
  • the corner feature corresponding to the target detection object includes at least a first corner feature and a second corner feature corresponding to the target object.
  • the size feature corresponding to the target detection object includes: the length feature and the width feature corresponding to the first corner feature of the target detection object, and the length feature and the width feature corresponding to the second corner feature of the target detection object.
  • the corner point may include a first corner point and a second corner point, and the first corner point and the second corner point may be a pair of opposite corner points.
  • the corner point feature may include the first corner point feature And the second corner point feature, in this way, the first corner point feature and the second corner point feature together represent the position information of the diagonal point of the image area where the target detection object is located, which can reduce the difficulty of distinguishing different target detection objects. happened.
  • the size feature corresponding to the target detection object can include the length feature and width feature corresponding to the first corner feature, and the length feature and width feature corresponding to the second corner feature of the target detection object. In this way, it can be detected according to the target. The size characteristics of the object at different corner features can further distinguish different target detection objects, so that the detection result of the target detection object is more accurate.
  • the first corner point may be the upper left corner point or the lower right corner point of the target detection object; correspondingly, the second corner point may be the upper right corner point or the lower left corner point corresponding to the target object.
  • the first corner point and the second corner point may be a pair of diagonal points, that is, the first corner point may be the upper left corner point, and the second corner point may be the lower right corner point.
  • the first corner point may be the upper right corner point
  • the second corner point may be the lower left corner point.
  • Step S13 based on the size feature and the corner point feature, extract the object feature corresponding to the target detection object from the image feature.
  • the image size corresponding to the target detection object can be determined according to the size feature corresponding to the target detection object, and the image position of the target detection object in the detection image can be determined according to the corner feature corresponding to the target detection object, so that Combining the size feature corresponding to the target detection object and the corner point feature can determine the object feature corresponding to the target detection object.
  • the object feature can characterize the feature of the target detection object in the image area of the detection image.
  • the object feature can indicate the target detection.
  • the image position of the object can be determined according to the size feature corresponding to the target detection object, and the image position of the target detection object in the detection image can be determined according to the corner feature corresponding to the target detection object, so that Combining the size feature corresponding to the target detection object and the corner point feature can determine the object feature corresponding to the target detection object.
  • the object feature can characterize the feature of the target detection object in the image area of the detection image.
  • the object feature can indicate the target detection.
  • the image position of the object can be
  • the feature area that has a mapping relationship with the image area of the target detection object in the detection image can be determined according to the size feature and the corner point feature, and then the feature area in the image The feature area of the feature extracts the object feature corresponding to the target detection object.
  • the image feature of the detection image can be obtained, and the image feature can be expressed as a feature map.
  • the feature area indicated by the size feature and the corner feature can be determined in the feature map, and the feature area corresponds to the image area of the target detection object in the detection image. Extracting the image feature of the feature area in the feature map can be used as the object feature corresponding to the target detection object. According to the characteristics of the object, the image area of the target detection object in the detection image can be determined.
  • the detection frame of the target detection object in the detection image can be determined based on the length feature and width feature corresponding to the first corner point feature and the length feature and width feature corresponding to the second corner point feature. , And then determine the intersection ratio between any two overlapping detection frames, and when the intersection ratio is greater than the preset threshold, merge any two overlapping detection frames into one detection frame.
  • the length feature and width feature corresponding to the first corner point feature can indicate a feature area in the feature map
  • the length feature and width feature corresponding to the second corner feature can indicate a feature in the feature map.
  • Area, the characteristic area indicated by the first corner point feature and the characteristic area indicated by the second corner point characteristic may be the same characteristic area, or may be different characteristic areas.
  • the detection frame can be used to frame the image area of the target detection object in the detection image in the detection image.
  • the detection frame can be a closed figure, for example , Square, rectangular and other quadrilaterals.
  • the detection frame can identify that the target detection object is in the image area of the detection image
  • the aforementioned corner point feature can indicate the positions of two diagonal points of the detection frame
  • the aforementioned size feature can indicate the length and width of the detection frame.
  • the same target detection object may have multiple detection frames, and multiple detection frames may overlap. Therefore, when the detection frames of the target detection object overlap, the difference between any two overlapping detection frames can be calculated. Intersection ratio, if the calculated intersection ratio is greater than the preset threshold, it can be considered that the two overlapping detection frames identify the same target detection object, and the larger detection frame of the two overlapping detection frames can be used as the target detection The detection frame of the object, delete the small detection frame.
  • a new detection frame can be synthesized based on two mutually overlapping detection frames, and the new detection frame is used as the detection frame of the target detection object, and the new detection frame includes the two detection frames before merging. In this way, the obtained detection frames can be further screened, and the detection frames of the same target detection object can be combined, so that one target detection object corresponds to one detection frame.
  • Fig. 2 shows a block diagram of a detection frame of a target detection object according to an embodiment of the present disclosure.
  • the detection frame as shown in FIG. 2 can be formed.
  • Step S14 Determine the category of the target detection object based on the characteristics of the object.
  • a neural network can be used to further perform feature feature extraction on the extracted object features, for example, convolution processing, normalization processing, etc. can be performed on the object features to obtain the target detection object category, for example, the target
  • the detection objects belong to categories such as vehicles, pedestrians, buildings, and public facilities. In this way, the category of the target detection object can be obtained from the object characteristics, and the target detection of the target detection object in the detection image can be realized.
  • the embodiment of the present disclosure determines the size feature and corner point feature corresponding to the target detection object by detecting the image feature of the image, and then extracts the object feature corresponding to the target detection object from the image feature based on the size feature and the corner point feature, and further based on the extracted object
  • the feature determines the category of the target detection object, so that the object characteristics of the target detection object and the classification of the target detection object are non-parallel. When the target detection object is classified, the object characteristics of the target detection object can be considered to obtain the classification result More accurate and improve the accuracy of target detection.
  • At least one level of convolution processing may be performed on the object feature to obtain the probability that the target detection object belongs to at least one preset category, and then according to the target detection object belonging to at least one preset category The probability of the category is to determine the category of the target detection object from the preset category.
  • the neural network can be used to further perform at least one level of convolution processing on the extracted object features, and the probability that the target detection object belongs to at least one predicted category can be obtained.
  • the preset category is pedestrian, vehicle, building, etc.
  • the probability that the target detection object belongs to multiple preset categories among pedestrians, vehicles, and buildings can be obtained. Then, the preset category with the highest probability can be determined as the category of the target detection object.
  • the object characteristics corresponding to the target detection object can be determined first, and then the target detection object can be used to classify the target detection object, and the category of the target detection object can be determined. In this way, a higher accuracy detection can be obtained. result.
  • a neural network can be used to obtain the category of the target detection object. The process of obtaining the category of the target detection object by using the neural network will be described below.
  • At least one level of convolution processing may be performed on the detection image to obtain the image characteristics of the detection image, and then corner pooling processing is performed on the image characteristics of the detection image to obtain target detection The corresponding size feature and corner feature of the object.
  • the neural network can include a multi-level convolutional layer and a corner pooling layer.
  • the detection image can be used as the input of the neural network, and then the neural network can be used to perform multi-level convolution processing on the detection image to obtain the detection image.
  • the corner pooling layer of the neural network is used to perform corner pooling processing on the image features of the detected image, and the size features and corner features corresponding to the target detection object can be obtained.
  • the convolution processing includes up-sampling processing and down-sampling processing;
  • the performing at least one level of convolution processing on the detection image to obtain the image characteristics of the detection image may include: Perform at least one level of down-sampling processing on the detected image to obtain a first feature map after at least one level of down-sampling processing; and obtain at least one level of down-sampling processing based on the first feature map after at least one level of up-sampling processing After the second feature map; based on the first feature map after the at least one level of down-sampling processing and the second feature map after the at least one level of up-sampling processing, the image feature of the detection image is obtained.
  • the convolution processing can include up-sampling processing and down-sampling processing.
  • a neural network can be used to perform multi-level down-sampling processing on the detection image to obtain the first feature map after each level of down-sampling processing.
  • the first feature map obtained after the last downsampling process in the level downsampling process is subjected to multi-level upsampling processing, and the second feature map obtained after each level of upsampling process can be obtained.
  • the image features of the detection image can be obtained according to the first feature map after multi-level down-sampling processing and the second feature map after multi-level up-sampling processing.
  • the first feature map after multi-level down-sampling processing can be combined with multiple The second feature map after the up-sampling process is subjected to feature fusion to obtain the image feature of the detection image.
  • the up-sampling process can be performed using the bilinear difference method, so that a more accurate second feature map can be obtained.
  • each level of the down-sampling process outputs a first feature image
  • each level of the up-sampling process outputs a second feature image, for the first feature image in the at least one level of up-sampling process
  • the second feature map output after the upper level of upsampling processing of the Nth level of upsampling processing is matched to the Nth level
  • the first feature map of the second feature map output after the sampling process is used as the input of the Nth upsampling process to obtain the second feature map output by the Nth upsampling process, where N is greater than 1. Positive integer.
  • the multi-level down-sampling process on the detection image can obtain the first feature map after each level down-sampling processing.
  • you can Up-sampling the first feature map by the first-level up-sampling process in the multi-level up-sampling process is used to obtain the second feature map after the first-level up-sampling process, and then can be based on the first-level up-sampling process
  • Two feature maps and the first feature map matching the second feature map to obtain the input of the second-level upsampling process for example, the second feature map is merged with the first feature map to obtain the second-level upsampling
  • the processed input, or the first feature map is subjected to convolution processing and then fused with the second feature map to obtain the input of the second level of upsampling processing.
  • the first feature map matching the second feature map may be the first feature map matching the image size of the second feature map.
  • the input is up-sampled by the second-level up-sampling process, and the second feature map after the second-level up-sampling process can be obtained.
  • the input of the third upsampling process is obtained, and so on, the Nth upsampling process can be obtained
  • N is a positive integer greater than 1. In this way, in the up-sampling process, the image features obtained by the down-sampling process can be considered, so that the extracted image features are more accurate.
  • corner pooling processing can be performed on the image features of the detection image to obtain the processing result, and then the first branch network is used to perform convolution processing on the processing result to obtain the target detection object corresponding Using the second branch network to perform convolution processing on the processing result to obtain the corner feature corresponding to the target detection object.
  • the number of channels in the first branch network and the second branch network are different.
  • the aforementioned neural network may include two branch networks, namely, a first branch network and a second branch network.
  • the first branch network can be used to convolve the image features of the detection image to obtain the feature map of the first branch network.
  • the feature map can correspond to 4 channels, of which one The channel corresponds to the length feature of the first corner point, one channel corresponds to the width feature of the first corner point, one channel corresponds to the length feature of the second corner point, and one channel corresponds to the width feature of the second corner point.
  • the second branch network can be used to perform convolution processing on the image features of the detected image to obtain a feature map of the second branch network.
  • the feature map can correspond to two channels, where one channel corresponds to the first corner feature. Indicate the position of the first corner point in the detection image, and the other channel corresponds to the second corner point feature, which can indicate the position of the second corner point in the detection image.
  • the image area where the target detection object is located can be determined according to the size feature and corner feature corresponding to the target detection object, and the possibility that different target detection objects cannot be distinguished can be reduced.
  • FIG. 3 shows a block diagram of obtaining a detection result of a target detection object by using a neural network according to an embodiment of the present disclosure.
  • the following describes the process of obtaining the category of the target detection object by using the above-mentioned neural network with an example.
  • the length and width of the detection image can be adjusted to an appropriate size without changing the aspect ratio of the detection image, for example, the width of the detection image can be adjusted to 800 pixels.
  • the neural network can include multi-level convolutional layers.
  • the neural network can be used to down-sampling the detection image. Each level of down-sampling processing can get a first feature map.
  • Four-level convolution processing can obtain four first feature maps of different sizes, which are respectively denoted as C 2 , C 3 , C 4 , and C 5 .
  • C 2 both length and width is twice the C 3
  • C 3 are the length and width of twice the C 4
  • C 4 are the length and width of twice the C 5.
  • C 5 can be passed through a 1*1 convolution kernel to obtain a new feature map F 5
  • F 5 has the same length and width as C 5 .
  • F 5 of multistage sampling process the sampling process on each stage may have a second characteristic of FIG. That is, the up-sampling process of double linear interpolation can be performed on F 5 to obtain a second feature map whose length and width are both magnified twice, which can be denoted as F′ 5 .
  • C 4 is calculated by a 1*1 convolution kernel to obtain a new feature map C′ 4 , C′ 4 and F′ 5 have the same size, and C′ 4 and F′ 5 are added together to obtain The input F 4 of the second upsampling process.
  • the second feature map F 2 output after the last-stage up-sampling processing can be obtained. F 2 length and width of the same C 2.
  • the second feature map F 2 can be subjected to corner pooling processing to obtain the processing result.
  • the processing result can go through the first branch network and the second branch network respectively.
  • Each branch network can include a 3*3 convolution kernel.
  • the first branch network can form a 4-channel feature map location
  • the second branch network can form A 2-channel feature map mask.
  • the two channels of the feature map mask respectively represent the upper left corner point feature and the lower right corner point feature
  • the four channels of the feature map location respectively represent the width feature dw and the length feature dh corresponding to the upper left corner point, and the width feature corresponding to the lower right corner point. dw and length characteristics dh.
  • the upper left corner and lower right corner feature point feature, the upper left corner corresponding to the width and length of features characteristic features and the width and length corresponding to the lower right corner feature, a feature region may be determined, F 2 extracts the second characteristic feature in FIG.
  • the image features of the area can be used to obtain the object features of the target detection object. For example, after the RoI Align layer, the corresponding image features in the feature area of the second feature map F 2 can be obtained, and then the 3*3 convolution kernel is used to perform the object features Classification, you can get the category of the target detection object in the detection image.
  • the detection frame of the target detection object can be obtained from the upper left corner point feature and the lower right corner point feature, as well as the dw and dh corresponding to the upper left corner point, and the dw and dh corresponding to the lower right corner point.
  • the width of the detection frame is calculated by the following formula (1):
  • w is the image width of the detection frame
  • s, ⁇ , and ⁇ can be mapping parameters, which can be obtained through network parameters
  • dw is the width feature.
  • non-maximum value suppression processing can be performed on the multiple detection frames of the target detection object, and the multiple detection frames of the target detection object can be merged into one to obtain the final target detection object.
  • the test results can be performed on the multiple detection frames of the target detection object, and the multiple detection frames of the target detection object can be merged into one to obtain the final target detection object.
  • the target detection solution provided by the embodiments of the present disclosure can more effectively predict the detection frame of the target detection object.
  • the detection frame is obtained through the corner points, thereby making the predicted detection frame more accurate and effectively alleviating the The problem of low accuracy of the predicted detection frame caused by overlap.
  • the prediction of the detection frame and the classification of the target detection object are performed non-parallel, that is, the size feature and corner feature indicating the position of the detection frame are first obtained, and then the target detection is performed by the object feature determined by the size feature and the corner feature Objects are classified, so that more accurate classification results can be obtained.
  • the present disclosure also provides target detection devices, electronic equipment, computer-readable storage media, and programs, all of which can be used to implement any target detection method provided in the present disclosure.
  • target detection devices electronic equipment, computer-readable storage media, and programs, all of which can be used to implement any target detection method provided in the present disclosure.
  • the writing order of the steps does not mean a strict execution order but constitutes any limitation on the implementation process.
  • the specific execution order of each step should be based on its function and possibility.
  • the inner logic is determined.
  • Fig. 4 shows a block diagram of a target detection device according to an embodiment of the present disclosure.
  • the target detection device includes:
  • the obtaining module 41 is used to obtain the detected image to be processed
  • the determining module 42 is configured to determine the size feature and corner point feature corresponding to the target detection object according to the image feature of the detection image;
  • the extraction module 43 is configured to extract the object feature corresponding to the target detection object from the image feature based on the size feature and the corner point feature;
  • the classification module 44 is configured to determine the category of the target detection object based on the characteristics of the object.
  • the determining module 42 is specifically configured to perform at least one level of convolution processing on the detection image to obtain the image characteristics of the detection image;
  • the point pooling process obtains the size feature and corner point feature corresponding to the target detection object.
  • the convolution processing includes up-sampling processing and down-sampling processing;
  • the determining module 42 is specifically configured to perform at least one level of down-sampling processing on the detection image to obtain at least one level A first feature map after downsampling processing; based on the first feature map after at least one level of upsampling processing, a second feature map after at least one level of down sampling processing is obtained; based on the at least one level of down sampling processing
  • the first feature map of and the second feature map after the at least one level of up-sampling processing are used to obtain the image feature of the detection image.
  • each level of the down-sampling process outputs a first feature image
  • each level of the up-sampling process outputs a second feature image
  • the determining module 42 is specifically configured to For the first-stage up-sampling process in the at least one-stage up-sampling process, the first feature map after the last-stage down-sampling process in the at least one-stage down-sampling process is used as the first-stage up-sampling process
  • the processed input; the second feature map output after the first-level upsampling processing is obtained; for the N-th up-sampling processing in the at least one-level up-sampling processing, the up-sampling processing of the N-th stage is up
  • the second feature map output after the first level upsampling processing and the first feature map matching the second feature map output after the Nth level upsampling processing are used as the input of the Nth level upsampling processing; to obtain the The second feature map output by the Nth upsampling process, where N is a
  • the determining module 42 is specifically configured to match the second feature map output after the upper-level upsampling processing of the Nth level of upsampling processing with the second feature map that matches the Nth level of upsampling processing.
  • Feature fusion is performed on the first feature map of the second feature map output after the upsampling processing to obtain the input of the Nth level of upsampling processing.
  • the determining module 42 is specifically configured to perform corner pooling processing on the image features of the detected image to obtain a processing result; and use the first branch network to roll up the processing result.
  • the extraction module 43 is specifically configured to determine, according to the size feature and the corner point feature, that there is a mapping relationship with the target detection object in the image area of the detection image The feature area of the image; extract the object feature corresponding to the target detection object in the feature area of the image feature.
  • the corner feature corresponding to the target detection object includes at least a first corner feature and a second corner feature corresponding to the target object;
  • the size feature corresponding to the target detection object includes: The length feature and width feature of the target detection object corresponding to the first corner feature, and the length feature and width feature of the target detection object corresponding to the second corner feature.
  • the device further includes: a merging module, configured to be based on the length feature and width feature corresponding to the first corner point feature and the length feature and width feature corresponding to the second corner point feature , Determine the detection frame of the target detection object in the detection image; determine the intersection ratio between any two overlapping detection frames; when the intersection ratio is greater than a preset threshold, combine the Any two overlapping detection frames are merged into one detection frame.
  • a merging module configured to be based on the length feature and width feature corresponding to the first corner point feature and the length feature and width feature corresponding to the second corner point feature , Determine the detection frame of the target detection object in the detection image; determine the intersection ratio between any two overlapping detection frames; when the intersection ratio is greater than a preset threshold, combine the Any two overlapping detection frames are merged into one detection frame.
  • the classification module 44 is specifically configured to perform at least one level of convolution processing on the object feature to obtain the probability that the target detection object belongs to each preset category; according to the target The detection object belongs to the probability of each preset category, and the category of the target detection object is determined from the preset category.
  • the functions or modules contained in the device provided in the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments.
  • the functions or modules contained in the device provided in the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments.
  • An embodiment of the present disclosure also provides an electronic device, including: a processor; a memory for storing executable instructions of the processor; wherein the processor is configured as the above method.
  • the electronic device can be provided as a terminal, server or other form of device.
  • Fig. 5 is a block diagram showing an electronic device 1900 according to an exemplary embodiment.
  • the electronic device 1900 may be provided as a server. 5
  • the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932, for storing instructions executable by the processing component 1922, such as application programs.
  • the application program stored in the memory 1932 may include one or more modules each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform the above-described methods.
  • the electronic device 1900 may also include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to the network, and an input output (I/O) interface 1958 .
  • the electronic device 1900 can operate based on an operating system stored in the memory 1932, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or the like.
  • a volatile computer-readable storage medium or a non-volatile computer-readable storage medium is also provided, such as the memory 1932 including computer program instructions, which can be processed by the electronic device 1900.
  • the component 1922 executes to complete the above method.
  • a computer program is also provided.
  • the computer program includes computer-readable code.
  • a processor in the electronic device executes to implement the foregoing method.
  • the present disclosure may be a system, method and/or computer program product.
  • the computer program product may include a computer-readable storage medium loaded with computer-readable program instructions for enabling a processor to implement various aspects of the present disclosure.
  • the computer-readable storage medium may be a tangible device that can hold and store instructions used by the instruction execution device.
  • the computer-readable storage medium may be, for example, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Non-exhaustive list of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM) Or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical encoding device, such as a printer with instructions stored thereon
  • RAM random access memory
  • ROM read-only memory
  • EPROM erasable programmable read-only memory
  • flash memory flash memory
  • SRAM static random access memory
  • CD-ROM compact disk read-only memory
  • DVD digital versatile disk
  • memory stick floppy disk
  • mechanical encoding device such as a printer with instructions stored thereon
  • the computer-readable storage medium used here is not interpreted as the instantaneous signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (for example, light pulses through fiber optic cables), or through wires Transmission of electrical signals.
  • the computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing/processing devices, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and/or a wireless network.
  • the network may include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers.
  • the network adapter card or network interface in each computing/processing device receives computer-readable program instructions from the network, and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing/processing device .
  • the computer program instructions used to perform the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or in one or more programming languages.
  • Source code or object code written in any combination, the programming language includes object-oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as "C" language or similar programming languages.
  • Computer-readable program instructions can be executed entirely on the user's computer, partly on the user's computer, executed as a stand-alone software package, partly on the user's computer and partly executed on a remote computer, or entirely on the remote computer or server carried out.
  • the remote computer can be connected to the user's computer through any kind of network-including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (for example, using an Internet service provider to connect to the user's computer) connection).
  • LAN local area network
  • WAN wide area network
  • an electronic circuit such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), can be customized by using the status information of the computer-readable program instructions.
  • FPGA field programmable gate array
  • PDA programmable logic array
  • the computer-readable program instructions are executed to realize various aspects of the present disclosure.
  • These computer-readable program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine that makes these instructions when executed by the processor of the computer or other programmable data processing device , A device that implements the functions/actions specified in one or more blocks in the flowcharts and/or block diagrams is produced. It is also possible to store these computer-readable program instructions in a computer-readable storage medium. These instructions make computers, programmable data processing apparatuses, and/or other devices work in a specific manner. Thus, the computer-readable medium storing the instructions includes An article of manufacture, which includes instructions for implementing various aspects of the functions/actions specified in one or more blocks in the flowcharts and/or block diagrams.
  • each block in the flowchart or block diagram may represent a module, program segment, or part of an instruction, and the module, program segment, or part of an instruction contains one or more components for realizing the specified logical function.
  • Executable instructions may also occur in a different order than the order marked in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, or they can sometimes be executed in the reverse order, depending on the functions involved.
  • each block in the block diagram and/or flowchart, and the combination of the blocks in the block diagram and/or flowchart can be implemented by a dedicated hardware-based system that performs the specified functions or actions Or it can be realized by a combination of dedicated hardware and computer instructions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Quality & Reliability (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Signal Processing For Digital Recording And Reproducing (AREA)
  • Storage Device Security (AREA)

Abstract

一种目标检测方法及装置、电子设备和存储介质，所述方法包括：获取待处理的检测图像（S11）；根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征（S12）；基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征（S13）；基于所述对象特征确定所述目标检测对象的类别（S14）。

Description

目标检测方法及装置、电子设备和存储介质
本公开要求在2019年10月29日提交中国专利局、申请号为201911038042.3、申请名称为“目标检测方法及装置、电子设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本公开中。
技术领域
本公开涉及计算机视觉技术领域，尤其涉及一种目标检测方法及装置、电子设备和存储介质。
背景技术
计算机视觉是利用计算机及相关设备对生物视觉进行模拟的一种技术，可以通过对采集的图像或视频进行处理，获得相应场景的三维信息。在计算机视觉的一个应用中，可以利用采集的图像或视频进行目标检测，确定目标对象类别以及在图像中的位置。
目前，目标检测技术可以利用神经网络直接确定目标对象的类别和定位的检测框。
发明内容
本公开提出了一种目标检测技术方案。
根据本公开的一方面，提供了一种目标检测方法，包括：获取待处理的检测图像；根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征；基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征；基于所述对象特征确定所述目标检测对象的类别。
在一种可能的实现方式中，所述根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征，包括：对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征；对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征。
在一种可能的实现方式中，所述卷积处理包括上采样处理和下采样处理；所述对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征，包括：对所述检测图像进行至少一级下采样处理，得到至少一级下采样处理后的第一特征图；基于所述至少一级上采样处理后的第一特征图，得到至少一级下采样处理后的第二特征图；基于所述至少一级下采样处理后的第一特征图和所述至少一级上采样处理后的第二特征图，得到所述检测图像的图像特征。
在一种可能的实现方式中，每一级所述下采样处理后输出一个第一特征图，每一级所述上采样处理后输出一个第二特征图像；所述基于所述至少一级上采样处理后的第一 特征图，得到至少一级下采样处理后的第二特征图，包括：针对所述至少一级上采样处理中的第一级上采样处理，将所述至少一级下采样处理中最后一级下采样处理后的第一特征图，作为所述第一级上采样处理的输入；得到所述第一级上采样处理后输出的第二特征图；针对所述至少一级上采样处理中的第N级上采样处理，将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入；得到所述第N级上采样处理输出的第二特征图，其中，N为大于1的正整数。
在一种可能的实现方式中，所述将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入，包括：将所述第N级上采样处理的上一级上采样处理后输出的第二特征图，与匹配于所述第N级上采样处理后输出的第二特征图的第一特征图进行特征融合，得到所述第N级上采样处理的输入。
在一种可能的实现方式中，所述对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征，包括：对所述检测图像的图像特征进行角点池化处理，得到处理结果；利用第一分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的尺寸特征；利用第二分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的角点特征；其中，第一分支网络与第二分支网络的通道数不同。
在一种可能的实现方式中，所述基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征，包括：根据所述尺寸特征和所述角点特征，确定与所述目标检测对象在所述检测图像中的图像区域存在映射关系的特征区域；在所述图像特征的特征区域提取所述目标检测对象对应的对象特征。
在一种可能的实现方式中，所述目标检测对象对应的角点特征包括至少所述目标对象对应的第一角点特征和第二角点特征；所述目标检测对象对应的尺寸特征包括：所述目标检测对象在第一角点特征对应的长度特征和宽度特征，以及所述目标检测对象在第二角点特征对应的长度特征和宽度特征。
在一种可能的实现方式中，所述方法还包括：基于所述第一角点特征对应的长度特征和宽度特征以及所述第二角点特征对应的长度特征和宽度特征，确定所述目标检测对象在所述检测图像中的检测框；确定任意两个相互重叠的检测框之间的交并比；在所述交并比大于预设阈值的情况下，将所述任意两个重叠的检测框合并为一个检测框。
在一种可能的实现方式中，其特征在于，所述基于所述对象特征确定所述目标检测 对象的类别，包括：对所述对象特征进行至少一级卷积处理，得到所述目标检测对象属于至少一个预设类别的概率；根据所述目标检测对象属于至少一个预设类别的概率，从所述预设类别中确定所述目标检测对象的类别。
根据本公开的一方面，提供了一种目标检测装置，包括：获取模块，用于获取待处理的检测图像；确定模块，用于根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征；提取模块，用于基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征；分类模块，用于基于所述对象特征确定所述目标检测对象的类别。
在一种可能的实现方式中，所述确定模块，具体用于对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征；对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征。
在一种可能的实现方式中，所述卷积处理包括上采样处理和下采样处理；所述确定模块，具体用于，对所述检测图像进行至少一级下采样处理，得到至少一级下采样处理后的第一特征图；基于所述至少一级上采样处理后的第一特征图，得到至少一级下采样处理后的第二特征图；基于所述至少一级下采样处理后的第一特征图和所述至少一级上采样处理后的第二特征图，得到所述检测图像的图像特征。
在一种可能的实现方式中，每一级所述下采样处理后输出一个第一特征图，每一级所述上采样处理后输出一个第二特征图像；所述确定模块，具体用于，针对所述至少一级上采样处理中的第一级上采样处理，将所述至少一级下采样处理中最后一级下采样处理后的第一特征图，作为所述第一级上采样处理的输入；得到所述第一级上采样处理后输出的第二特征图；针对所述至少一级上采样处理中的第N级上采样处理，将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入；得到所述第N级上采样处理输出的第二特征图，其中，N为大于1的正整数。
在一种可能的实现方式中，具体用于将所述第N级上采样处理的上一级上采样处理后输出的第二特征图，与匹配于所述第N级上采样处理后输出的第二特征图的第一特征图进行特征融合，得到所述第N级上采样处理的输入。
在一种可能的实现方式中，所述确定模块，具体用于，对所述检测图像的图像特征进行角点池化处理，得到处理结果；利用第一分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的尺寸特征；利用第二分支网络对所述处理结果进行卷积处理， 得到目标检测对象对应的角点特征；其中，第一分支网络与第二分支网络的通道数不同。
在一种可能的实现方式中，所述提取模块，具体用于，根据所述尺寸特征和所述角点特征，确定与所述目标检测对象在所述检测图像中的图像区域存在映射关系的特征区域；在所述图像特征的特征区域提取所述目标检测对象对应的对象特征。
在一种可能的实现方式中，所述目标检测对象对应的角点特征包括至少所述目标对象对应的第一角点特征和第二角点特征；所述目标检测对象对应的尺寸特征包括：所述目标检测对象在第一角点特征对应的长度特征和宽度特征，以及所述目标检测对象在第二角点特征对应的长度特征和宽度特征。
在一种可能的实现方式中，所述装置还包括：合并模块，用于基于所述第一角点特征对应的长度特征和宽度特征以及所述第二角点特征对应的长度特征和宽度特征，确定所述目标检测对象在所述检测图像中的检测框；确定任意两个相互重叠的检测框之间的交并比；在所述交并比大于预设阈值的情况下，将所述任意两个重叠的检测框合并为一个检测框。
在一种可能的实现方式中，所述分类模块，具体用于，对所述对象特征进行至少一级卷积处理，得到所述目标检测对象属于至少一个预设类别的概率；根据所述目标检测对象属于至少一个预设类别的概率，从所述预设类别中确定所述目标检测对象的类别。
根据本公开的一方面，提供了一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为：执行上述目标检测方法。
根据本公开的一方面，提供了一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述目标检测方法。
根据本公开的一方面，提供了一种计算机程序，其中，所述计算机程序包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现上述目标检测方法。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，而非限制本公开。
根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。
图1示出根据本公开实施例的目标检测方法的流程图。
图2示出根据本公开实施例的目标检测对象的检测框的框图。
图3示出根据本公开实施例的利用神经网络得到目标检测对象的类别的框图。
图4示出根据本公开实施例的目标检测装置的框图。
图5示出根据本公开实施例的电子设备示例的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合，例如，包括A、B、C中的至少一种，可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
另外，为了更好地说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
本公开实施例提供的目标检测方案，可以获取待处理的检测图像，然后根据检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征，然后基于尺寸特征和角点特征，可以从图像特征中提取目标检测对象对应的对象特征，再基于目标检测对象对应的对象特征，确定目标检测对象的类别，从而可以得到目标检测的检测结果。通过这种方式，可以先确定目标检测对象对应的对象特征，然后根据对应的对象特征，再对目标检测对象进行分类，从而可以在目标检测中，非并行确定目标检测对象的对象特征和目标对象的类别，使得得到的检测结果可以更加准确，可以提高检测结果的精度。
在相关技术中，目标检测通常需要密集采集形成初步检测框的锚点，但是大量的锚点中会存在很多无效的锚点，会耗费处理的时间以及存储空间。此外，在相关技术的目标检测过程中，目标检测对象的检测框和类别是并行进行确定的，从而在确定目标检测 对象的类别时无法考虑检测框的信息，导致得到的检测结果不够准确。本公开实施例提供的目标检测方案，通过确定目标检测对象的角点和尺寸可以先确定目标检测对象的对象特征，减少了由于采集大量的锚点而带来的时间和存储空间的浪费，并且，通过角点和尺寸得到的对象特征，可以在两个目标检测对象发现重合的情况下，减小通过由锚点确定的中心点区分不同的目标检测对象的难度，从而本公开实施例提供的目标检测方案，可以通过角点特征区分不同目标检测对象，可以节省采集锚点耗费时间和存储空间，提高得到检测结果的效率，并且得到的检测结果具有较高的准确性。
本公开实施例提供的信息处理方案，可以应用于任何需要进行目标检测的场景。例如，可以应用于对采集的视频进行目标检测的场景，根据检测结果可以得到视频中目标检测对象的轨迹。再例如，可以应用于安防场景中，根据检测结果可以对嫌疑人进行识别和追踪。下面通过实施例对本公开提供的目标检测方案进行说明。
图1示出根据本公开实施例的目标检测方法的流程图。该目标检测方法可以由终端设备、服务器或其它目标检测装置执行，其中，终端设备可以为用户设备(User Equipment，UE)、移动设备、用户终端、终端、蜂窝电话、无绳电话、个人数字处理(Personal Digital Assistant，PDA)、手持设备、计算设备、车载设备、可穿戴设备等。在一些可能的实现方式中，该目标检测方法可以通过处理器调用存储器中存储的计算机可读指令的方式来实现。下面以目标检测装置作为执行主体为例对本公开实施例的目标检测方法进行说明。
如图1所示，所述目标检测方法包括以下步骤：
步骤S11，获取待处理的检测图像。
在本公开实施例中，目标检测装置可以具有图像采集功能，可以对当前场景进行拍摄，得到待处理的检测图像。或者，目标检测装置可以由其他设备获取待处理的检测图像。检测图像可以是单独采集的图像，或者，可以是视频流中的一个图像帧。
这里，获取的待处理的检测图像可以是经过预处理的检测图像，例如，可以经过图像放缩、图像增强、图像滤波等预处理。例如，可以在不改变检测图像的长宽比的情况下，将检测图像的长和宽调整到合适的尺寸，得到经过预处理的检测图像。
步骤S12，根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征。
在本公开实施例中，可以利用神经网络提取检测图像的图像特征，由检测图像的图像特征可以确定目标检测对象对应的尺寸特征，以及，确定目标检测对象对应的角点特征。这里，目标检测对象可以是检测图像中需要进行检测的对象，例如，行人、车辆、 建筑物、标识等对象的图像，目标检测对象的尺寸特征可以表征目标检测对象所在的图像区域的尺寸特征，例如，在目标检测对象在检测图像的图像区域表示为四边形的情况下，目标检测对象对应的尺寸特征可以是四边形对应的长度特征和/或宽度特征。角点特征可以表征目标检测对象所在的图像区域的对角点的位置信息。
在一种可能的实现方式中，目标检测对象对应的角点特征包括至少所述目标对象对应的第一角点特征和第二角点特征。所述目标检测对象对应的尺寸特征包括：所述目标检测对象在第一角点特征对应的长度特征和宽度特征，以及所述目标检测对象在第二角点特征对应的长度特征和宽度特征。
在该实现方式中，角点可以包括第一角点和第二角点，第一角点和第二角点可以是一对对角点，相应地，角点特征可以包括第一角点特征和第二角点特征，这样，通过第一角点特征和第二角点特征共同表示目标检测对象所在的图像区域的对角点的位置信息，可以减小很难区分不同的目标检测对象问题的发生。相应地，目标检测对象对应的尺寸特征可以包括第一角点特征对应的长度特征和宽度特征，以及，目标检测对象在第二角点特征对应的长度特征和宽度特征，这样，可以根据目标检测对象在不同角点特征对应的尺寸特征，进一步区分不同的目标检测对象，使得针对目标检测对象的检测结果更加准确。
这里，第一角点可以是目标检测对象的左上角点或右下角点；相应地，第二角点可以是目标对象对应的右上角点或左下角点。第一角点和第二角点可以是对角点对，也就是说，第一角点可以是左上角点，第二角点可以是右下角点。或者，第一角点可以是右上角点，第二角点可以是左下角点。
步骤S13，基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征。
在本公开实施例中，可以根据目标检测对象对应的尺寸特征可以确定目标检测对象对应的图像尺寸，根据目标检测对象对应的角点特征，可以确定目标检测对象在检测图像中的图像位置，从而将目标检测对象对应的尺寸特征和角点特征相结合，可以确定目标检测对象对应的对象特征，该对象特征可以表征目标检测对象在检测图像中图像区域所对应特征，该对象特征可以指示目标检测对象的图像位置。
在一种可能的实现方式中，在提取目标检测对象对应的对象特征，可以根据尺寸特征和角点特征，确定与目标检测对象在检测图像中的图像区域存在映射关系的特征区域，然后在图像特征的特征区域提取目标检测对象对应的对象特征。
在该实现方式中，对检测图像进行特征提取，可以得到检测图像的图像特征，该图像特征可以表示为特征图。根据目标检测对象对应的尺寸特征和角点特征，可以在该特征图中确定由尺寸特征和角点特征指示的特征区域，该特征区域对应目标检测对象在检测图像中的图像区域。在特征图中提取该特征区域的图像特征，可以作为目标检测对象对应对象特征。根据对象特征可以确定目标检测对象在检测图像中的图像区域。
在一种可能的实现方式中，可以基于第一角点特征对应的长度特征和宽度特征以及第二角点特征对应的长度特征和宽度特征，确定目标检测对象在所述检测图像中的检测框，然后确定任意两个相互重叠的检测框之间的交并比，在交并比大于预设阈值的情况下，将任意两个重叠的检测框合并为一个检测框。
在该实现方式中，第一角点特征对应的长度特征和宽度特征可以在上述特征图中指示一个特征区域，第二角点特征对应的长度特征和宽度特征可以在上述特征图中指示一个特征区域，第一角点特征指示的特征区域和第二角点特征指示的特征区域可以是同一个特征区域，也可以是不同的特征区域。根据特征区域与目标检测对象在检测图像中的图像区域之间的映射关系，可以利用检测框在检测图像中框出目标检测对象在检测图像中的图像区域，该检测框可以是封闭图形，例如，正方形、长方形等四边形。检测框可以标识目标检测对象在检测图像的图像区域，上述角点特征可以指示检测框的两个对角点的位置，上述尺寸特征可以指示检测框的长度和宽度。
这里，同一个目标检测对象可能存在多个检测框，多个检测框之间可能发生重叠，从而在目标检测对象的检测框重叠的情况下，可以计算任意两个相互重叠的检测框之间的交并比，如果计算的交并比大于预设阈值，则可以认为这两个重叠的检测框标识的是同一个目标检测对象，可以将两个重叠的检测框中大的检测框作为目标检测对象的检测框，删除其中小的检测框。或者，可以根据两个相互重叠的检测框合成一个新的检测框，将新的检测框作为目标检测对象的检测框，新的检测框包括合并前的两个检测框。这样，可以对得到的检测框进一步筛选，合并同一个目标检测对象的检测框，使一个目标检测对象对应一个检测框。
图2示出根据本公开实施例的目标检测对象的检测框的框图。以第一角点为左上角点为例，根据第一角点特征、第一角点特征对应的长度特征和宽度特征，可以形成如图2所示的检测框。
步骤S14，基于所述对象特征确定所述目标检测对象的类别。
在本公开实施例中，可以利用神经网络对提取的对象特征进一步进行特征特征提取， 例如，对对象特征进行卷积处理、归一化处理等，可以得到目标检测对象的类别，例如，该目标检测对象属于车辆、行人、建筑物、公共设施等类别。通过这种方式，可以由对象特征得到目标检测对象的类别，实现对检测图像中的目标检测对象进行目标检测。
本公开实施例通过检测图像的图像特征确定目标检测对象对应的尺寸特征和角点特征，然后基于尺寸特征和角点特征，从图像特征中提取目标检测对象对应的对象特征，进一步基于提取的对象特征确定目标检测对象的类别，从而确定目标检测对象的对象特征和对目标检测对象进行分类是非并行进行的，在对目标检测对象进行分类是可以考虑目标检测对象的对象特征，从而得到的分类结果更加准确，提高目标检测的精度。
在一种可能的实现方式中，可以对所述对象特征进行至少一级卷积处理，得到所述目标检测对象属于至少一个预设类别的概率，再根据所述目标检测对象属于至少一个预设类别的概率，从所述预设类别中确定所述目标检测对象的类别。
在该实现方式中，可以利用神经网络进一步对提取的对象特征进行至少一级卷积处理，可以得到目标检测对象属于至少一个预测类别的概率，例如，预设类别是行人、车辆、建筑物等任意一个类别，经过对对象特征进一步进行卷积处理，可以得到目标检测对象分别属于行人、车辆、建筑物中多个预设类别的概率。然后可以将概率最高的预设类别确定为目标检测对象的类别。
通过本公开实施例提供的目标检测方案，可以先确定目标检测对象对应的对象特征，然后再利用对象特征对目标检测对象进行分类，确定目标检测对象的类别，这样，可以得到准确性较高检测结果。本公开实施例提供的目标检测方案，可以利用神经网络得到目标检测对象的类别，下面对利用神经网络得到目标检测对象的类别的过程进行说明。
在一个可能的实现方式中，可以对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征，然后对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征。
在该实现方式中，神经网络可以包括多级卷积层和角点池化层，可以将检测图像作为神经网络的输入，然后利用神经网络对检测图像进行多级卷积处理，可以得到检测图像的图像特征，然后再利用神经网络的角点池化层对检测图像的图像特征进行角点池化处理，可以得到目标检测对象对应的尺寸特征和角点特征。
在该实现方式的一个示例中，所述卷积处理包括上采样处理和下采样处理；所述对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征，可以包括：对所述检测图像进行至少一级下采样处理，得到至少一级下采样处理后的第一特征图；基于 所述至少一级上采样处理后的第一特征图，得到至少一级下采样处理后的第二特征图；基于所述至少一级下采样处理后的第一特征图和所述至少一级上采样处理后的第二特征图，得到所述检测图像的图像特征。
在该示例中，卷积处理可以包括上采样处理和下采样处理，可以先利用神经网络对检测图像进行多级下采样处理，得到每级下采样处理后的第一特征图，然后再对多级下采样处理中最后一级下采样处理后得到的第一特征图进行多级上采样处理，可以得到每级上采样处理后得到的第二特征图。然后可以根据多级下采样处理后的第一特征图与多级上采样处理后的第二特征图得到检测图像的图像特征，例如，可以将多级下采样处理后的第一特征图与多级上采样处理后的第二特征图进行特征融合，得到检测图像的图像特征。这里，可以利用双线性差值方式进行上采样处理，从而可以得到比较准确的第二特征图。
在该示例中，每一级所述下采样处理后输出一个第一特征图，每一级所述上采样处理后输出一个第二特征图像，针对所述至少一级上采样处理中的第一级上采样处理，将所述至少一级下采样处理中最后一级下采样处理后的第一特征图，作为所述第一级上采样处理的输入，得到所述第一级上采样处理后输出的第二特征图。针对所述至少一级上采样处理中的第N级上采样处理，将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入，得到所述第N级上采样处理输出的第二特征图，其中，N为大于1的正整数。
在该示例中，对检测图像进行多级下采样处理可以得到每级下采样处理后的第一特征图，针对多级下采样处理中最后一级下采样处理后得到的第一特征图，可以利用多级上采样处理中的第一级上采样处理对该第一特征图进行上采样，得到第一级上采样处理后的第二特征图，然后可以根据第一级上采样处理后的第二特征图以及与该第二特征图匹配的第一特征图，得到第二级上采样处理的输入，例如，将该第二特征图与该第一特征图进行融合，得到第二级上采样处理的输入，或者，对该第一特征图进行卷积处理后与该第二特征图进行融合，得到第二级上采样处理的输入。这里与该第二特征图匹配的第一特征图可以是与该第二特征图的图像尺寸匹配的第一特征图。利用第二级上采样处理对输入进行上采样，可以得到第二级上采样处理后的第二特征图。然后根据第二级上采样处理后的第二特征图以及与该第二特征图匹配的第一特征图，得到第三级上采样处理的输入，依次类推，可以得到第N级上采样处理后的第二特征图。其中，N是大于1的 正整数。这样，在上采样处理过程中，可以考虑下采样处理得到的图像特征，从而使得提取的图像特征更加准确。
在该实现方式的一个示例中，可以对所述检测图像的图像特征进行角点池化处理，得到处理结果，然后利用第一分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的尺寸特征，利用第二分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的角点特征。第一分支网络与第二分支网络的通道数不同。
在该示例中，上述神经网络可以包括两个分支网络，即，第一分支网络和第二分支网络。在利用神经网络提取检测图像的图像特征之后，可以利用第一分支网络对检测图像的图像特征进行卷积处理，得到第一分支网络的特征图，该特征图可以对应4个通道，其中，一个通道对应第一角点的长度特征，一个通道对应第一角点的宽度特征，一个通道对应第二角点的长度特征，一个通道对应第二角点的宽度特征。相应地，可以利用第二分支网络对检测图像的图像特征进行卷积处理，得到第二分支网络的特征图，该特征图可以对应2个通道，其中，一个通道对应第一角点特征，可以指示第一角点在检测图像中的位置，另一个通道对应第二角点特征，可以指示第二角点在检测图像中的位置。这样，可以根据目标检测对象对应的尺寸特征和角点特征，可以确定目标检测对象所在的图像区域，并且，可以减小无法区分不同的目标检测对象的可能性。
图3示出根据本公开实施例的利用神经网络得到目标检测对象的检测结果的框图。
下面结合一示例对利用上述神经网络得到目标检测对象的类别的过程进行说明。可以在不改变检测图像的长宽比的情况下，将检测图像的长和宽调整到合适的尺寸，例如，将检测图像的宽调整到800像素。然后将调整后的检测图像输入到神经网络中，神经网络可以包括多级卷积层，可以先利用神经网络对检测图像进行下采样处理，每级下采样处理可以得到一个第一特征图，进行4级卷积处理，可以得到4个不同尺寸的第一特征图，分别记为C 2,C 3,C 4,C 5。其中，C 2的长和宽均是C 3的两倍，C 3的长和宽均是C 4的两倍，C 4的长和宽均是C 5的两倍。然后可以将C 5经过一个1*1的卷积核，得到一个新的特征图F 5，F 5长和宽与C 5相同。对F 5进行多级上采样处理，每级上采样处理可以得第二特征图。即，可以对F 5进行双线形插值的上采样处理，得到一个长和宽都放大两倍的第二特征图，可以记为F′ 5。C 4经过一个1*1的卷积核计算得到一个新的特征图C′ 4，C′ 4与F′ 5的尺寸相同，将C′ 4与F′ 5两个特征图相加，可以得到第二级上采样处理的输入F 4。对F 4进行上采样处理，得到一个长和宽都放大两倍的第二特征图F′ 4，然后将C 3经过一个1*1的卷积核计算得到一个新的特征图C′ 3，C′ 3与F′ 4的尺寸相同，将C′ 3与F′ 4两个特征图相加，可以得到第三级上采样处 理的输入F 3。以此类推，经过多次上采样处理，可以得到最后一级上采样处理后输出的第二特征图F 2。F 2的长和宽与C 2相同。
然后可以将第二特征图F 2进行角点池化处理，得到处理结果。该处理结果可以分别经过第一分支网络和第二分支网络，每个分支网络可以包括3*3的卷积核，然后第一分支网络可以形成4通道的特征图location，第二分支网络可以形成一个2通道的特征图mask。其中，特征图mask的两通道分别表示左上角点特征和右下角点特征，特征图location的四通道分别表示左上角点对应的宽度特征dw和长度特征dh，以及，右下角点对应的宽度特征dw和长度特征dh。
根据左上角点特征和右下角点特征、左上角点对应的宽度特征和长度特征以及右下角点对应的宽度特征和长度特征，可以确定一个特征区域，在第二特征图F 2中提取该特征区域的图像特征，可以得到目标检测对象的对象特征，例如经过RoI Align层，可以得到在第二特征图F 2在特征区域内对应的图像特征，然后利用3*3的卷积核对对象特征进行分类，可以得到检测图像中目标检测对象的类别。
这里，可以由左上角点特征和右下角点特征，以及左上角点对应的dw和dh、右下角点对应的dw和dh得到目标检测对象的检测框。
以检测框的宽度为例，检测框的宽度通过下述公式(1)计算得到：
w＝s*β*e dw*α     公式(1)；
其中，w是检测框的图像宽度；s、α和β可以是映射参数，可通过网络参数获到；dw是宽度特征。
在目标检测对象的检测框为多个的情况下，可以对目标检测对象的多个检测框进行非极大值抑制处理，将目标检测对象的多个检测框合并为一个，得到目标检测对象最终的检测结果。
本公开实施例提供的目标检测方案，可以更加有效的预测目标检测对象的检测框，该检测框是通过角点得到的，从而可以使得预测的检测框更加准确，有效地缓解了由于目标检测对象重叠带来的预测的检测框精度低的问题。此外，对于检测框的预测和目标检测对象的分类是非并行进行的，即，先得到指示检测框位置的尺寸特征和角点特征，然后再通过尺寸特征和角点特征确定的对象特征对目标检测对象进行分类，从而可以得到更加准确的分类结果。
可以理解，本公开提及的上述各个方法实施例，在不违背原理逻辑的情况下，均可 以彼此相互结合形成结合后的实施例，限于篇幅，本公开不再赘述。
此外，本公开还提供了目标检测装置、电子设备、计算机可读存储介质、程序，上述均可用来实现本公开提供的任一种目标检测方法，相应技术方案和描述和参见方法部分的相应记载，不再赘述。
本领域技术人员可以理解，在具体实施方式的上述方法中，各步骤的撰写顺序并不意味着严格的执行顺序而对实施过程构成任何限定，各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
图4示出根据本公开实施例的目标检测装置的框图，如图4所示，所述目标检测装置包括：
获取模块41，用于获取待处理的检测图像；
确定模块42，用于根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征；
提取模块43，用于基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征；
分类模块44，用于基于所述对象特征确定所述目标检测对象的类别。
在一种可能的实现方式中，所述确定模块42，具体用于对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征；对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征。
在一种可能的实现方式中，所述卷积处理包括上采样处理和下采样处理；所述确定模块42，具体用于，对所述检测图像进行至少一级下采样处理，得到至少一级下采样处理后的第一特征图；基于所述至少一级上采样处理后的第一特征图，得到至少一级下采样处理后的第二特征图；基于所述至少一级下采样处理后的第一特征图和所述至少一级上采样处理后的第二特征图，得到所述检测图像的图像特征。
在一种可能的实现方式中，每一级所述下采样处理后输出一个第一特征图，每一级所述上采样处理后输出一个第二特征图像；所述确定模块42，具体用于，针对所述至少一级上采样处理中的第一级上采样处理，将所述至少一级下采样处理中最后一级下采样处理后的第一特征图，作为所述第一级上采样处理的输入；得到所述第一级上采样处理后输出的第二特征图；针对所述至少一级上采样处理中的第N级上采样处理，将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入；得到所述第N 级上采样处理输出的第二特征图，其中，N为大于1的正整数。
在一种可能的实现方式中，所述确定模块42，具体用于将所述第N级上采样处理的上一级上采样处理后输出的第二特征图，与匹配于所述第N级上采样处理后输出的第二特征图的第一特征图进行特征融合，得到所述第N级上采样处理的输入。
在一种可能的实现方式中，所述确定模块42，具体用于，对所述检测图像的图像特征进行角点池化处理，得到处理结果；利用第一分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的尺寸特征；利用第二分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的角点特征；其中，第一分支网络与第二分支网络的通道数不同。
在一种可能的实现方式中，所述提取模块43，具体用于，根据所述尺寸特征和所述角点特征，确定与所述目标检测对象在所述检测图像中的图像区域存在映射关系的特征区域；在所述图像特征的特征区域提取所述目标检测对象对应的对象特征。
在一种可能的实现方式中，所述目标检测对象对应的角点特征包括至少所述目标对象对应的第一角点特征和第二角点特征；所述目标检测对象对应的尺寸特征包括：所述目标检测对象在第一角点特征对应的长度特征和宽度特征，以及所述目标检测对象在第二角点特征对应的长度特征和宽度特征。
在一种可能的实现方式中，所述装置还包括：合并模块，用于基于所述第一角点特征对应的长度特征和宽度特征以及所述第二角点特征对应的长度特征和宽度特征，确定所述目标检测对象在所述检测图像中的检测框；确定任意两个相互重叠的检测框之间的交并比；在所述交并比大于预设阈值的情况下，将所述任意两个重叠的检测框合并为一个检测框。
在一种可能的实现方式中，所述分类模块44，具体用于，对所述对象特征进行至少一级卷积处理，得到所述目标检测对象属于各个预设类别的概率；根据所述目标检测对象属于各个预设类别的概率，从所述预设类别中确定所述目标检测对象的类别。
在一些实施例中，本公开实施例提供的装置具有的功能或包含的模块可以用于执行上文方法实施例描述的方法，其具体实现可以参照上文方法实施例的描述，为了简洁，这里不再赘述。
本公开实施例还提出一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为上述方法。
电子设备可以被提供为终端、服务器或其它形态的设备。
图5是根据一示例性实施例示出的一种电子设备1900的框图。例如，电子设备1900可 以被提供为一服务器。参照图5，电子设备1900包括处理组件1922，其进一步包括一个或多个处理器，以及由存储器1932所代表的存储器资源，用于存储可由处理组件1922的执行的指令，例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理组件1922被配置为执行指令，以执行上述方法。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理，一个有线或无线网络接口1950被配置为将电子设备1900连接到网络，和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作***，例如Windows ServerTM，Mac OS XTM，UnixTM,LinuxTM，FreeBSDTM或类似。
在示例性实施例中，还提供了一种易失性计算机可读存储介质或非易失性计算机可读存储介质，例如包括计算机程序指令的存储器1932，上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述方法。
在示例性实施例中，还提供了一种计算机程序，计算机程序包括计算机可读代码，当计算机可读代码在电子设备中运行时，电子设备中的处理器执行用于实现上述方法。
本公开可以是***、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处 理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可读程序指令，从而实现本公开的各个方面。
这里参照根据本公开实施例的方法、装置(***)和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生产出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的***、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或 多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的***来实现，或者可以用专用硬件与计算机指令的组合来实现。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中技术的技术改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (23)

  1. 一种目标检测方法，包括：
    获取待处理的检测图像；
    根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征；
    基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征；
    基于所述对象特征确定所述目标检测对象的类别。
  2. 根据权利要求1所述的方法，其特征在于，所述根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征，包括：
    对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征；
    对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征。
  3. 根据权利要求2所述的方法，其特征在于，所述卷积处理包括上采样处理和下采样处理；所述对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征，包括：
    对所述检测图像进行至少一级下采样处理，得到至少一级下采样处理后的第一特征图；
    基于所述至少一级上采样处理后的第一特征图，得到至少一级下采样处理后的第二特征图；
    基于所述至少一级下采样处理后的第一特征图和所述至少一级上采样处理后的第二特征图，得到所述检测图像的图像特征。
  4. 根据权利要求3所述的方法，其特征在于，每一级所述下采样处理后输出一个第一特征图，每一级所述上采样处理后输出一个第二特征图像；
    所述基于所述至少一级上采样处理后的第一特征图，得到至少一级下采样处理后的第二特征图，包括：
    针对所述至少一级上采样处理中的第一级上采样处理，将所述至少一级下采样处理中最后一级下采样处理后的第一特征图，作为所述第一级上采样处理的输入；
    得到所述第一级上采样处理后输出的第二特征图；
    针对所述至少一级上采样处理中的第N级上采样处理，将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入；
    得到所述第N级上采样处理输出的第二特征图，其中，N为大于1的正整数。
  5. 根据权利要求4所述的方法，其特征在于，所述将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入，包括：
    将所述第N级上采样处理的上一级上采样处理后输出的第二特征图，与匹配于所述第N级上采样处理后输出的第二特征图的第一特征图进行特征融合，得到所述第N级上采样处理的输入。
  6. 根据权利要求2所述的方法，其特征在于，所述对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征，包括：
    对所述检测图像的图像特征进行角点池化处理，得到处理结果；
    利用第一分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的尺寸特征；
    利用第二分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的角点特征；其中，第一分支网络与第二分支网络的通道数不同。
  7. 根据权利要求1至6任意一项所述的方法，其特征在于，所述基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征，包括：
    根据所述尺寸特征和所述角点特征，确定与所述目标检测对象在所述检测图像中的图像区域存在映射关系的特征区域；
    在所述图像特征的特征区域提取所述目标检测对象对应的对象特征。
  8. 根据权利要求7所述的方法，其特征在于，所述目标检测对象对应的角点特征包括至少所述目标对象对应的第一角点特征和第二角点特征；
    所述目标检测对象对应的尺寸特征包括：所述目标检测对象在第一角点特征对应的长度特征和宽度特征，以及所述目标检测对象在第二角点特征对应的长度特征和宽度特征。
  9. 根据权利要求8所述的方法，其特征在于，所述方法还包括：
    基于所述第一角点特征对应的长度特征和宽度特征以及所述第二角点特征对应的长度特征和宽度特征，确定所述目标检测对象在所述检测图像中的检测框；
    确定任意两个相互重叠的检测框之间的交并比；
    在所述交并比大于预设阈值的情况下，将所述任意两个重叠的检测框合并为一个检测框。
  10. 根据权利要求1至9任意一项所述的方法，其特征在于，所述基于所述对象特征确定所述目标检测对象的类别，包括：
    对所述对象特征进行至少一级卷积处理，得到所述目标检测对象属于至少一个预设类别的概率；
    根据所述目标检测对象属于至少一个预设类别的概率，从所述预设类别中确定所述目标检测对象的类别。
  11. 一种目标检测装置，包括：
    获取模块，用于获取待处理的检测图像；
    确定模块，用于根据所述检测图像的图像特征，确定目标检测对象对应的尺寸特征和角点特征；
    提取模块，用于基于所述尺寸特征和所述角点特征，从所述图像特征中提取所述目标检测对象对应的对象特征；
    分类模块，用于基于所述对象特征确定所述目标检测对象的类别。
  12. 根据权利要求11所述的装置，其特征在于，所述确定模块，具体用于对所述检测图像进行至少一级卷积处理，得到所述检测图像的图像特征；对所述检测图像的图像特征进行角点池化处理，得到目标检测对象对应的尺寸特征和角点特征。
  13. 根据权利要求12所述的装置，其特征在于，所述卷积处理包括上采样处理和下采样处理；所述确定模块，具体用于，
    对所述检测图像进行至少一级下采样处理，得到至少一级下采样处理后的第一特征图；
    基于所述至少一级上采样处理后的第一特征图，得到至少一级下采样处理后的第二特征图；
    基于所述至少一级下采样处理后的第一特征图和所述至少一级上采样处理后的第二特征图，得到所述检测图像的图像特征。
  14. 根据权利要求13所述的装置，其特征在于，每一级所述下采样处理后输出一个第一特征图，每一级所述上采样处理后输出一个第二特征图像；所述确定模块，具体用于，
    针对所述至少一级上采样处理中的第一级上采样处理，将所述至少一级下采样处理中最后一级下采样处理后的第一特征图，作为所述第一级上采样处理的输入；
    得到所述第一级上采样处理后输出的第二特征图；
    针对所述至少一级上采样处理中的第N级上采样处理，将所述第N级上采样处理的上一级上采样处理后输出的第二特征图以及匹配于所述第N级上采样处理后输出的第二特征图的第一特征图，作为所述第N级上采样处理的输入；
    得到所述第N级上采样处理输出的第二特征图，其中，N为大于1的正整数。
  15. 根据权利要求14所述的装置，其特征在于，所述确定模块，具体用于将所述第N级上采样处理的上一级上采样处理后输出的第二特征图，与匹配于所述第N级上采样处理后输出的第二特征图的第一特征图进行特征融合，得到所述第N级上采样处理的输入。
  16. 根据权利要求12所述的装置，其特征在于，所述确定模块，具体用于，
    对所述检测图像的图像特征进行角点池化处理，得到处理结果；
    利用第一分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的尺寸特征；
    利用第二分支网络对所述处理结果进行卷积处理，得到目标检测对象对应的角点特征；其中，第一分支网络与第二分支网络的通道数不同。
  17. 根据权利要求11至16任意一项所述的装置，其特征在于，所述提取模块，具体用于，
    根据所述尺寸特征和所述角点特征，确定与所述目标检测对象在所述检测图像中的图像区域存在映射关系的特征区域；
    在所述图像特征的特征区域提取所述目标检测对象对应的对象特征。
  18. 根据权利要求17所述的装置，其特征在于，所述目标检测对象对应的角点特征包括至少所述目标对象对应的第一角点特征和第二角点特征；
    所述目标检测对象对应的尺寸特征包括：所述目标检测对象在第一角点特征对应的长度特征和宽度特征，以及所述目标检测对象在第二角点特征对应的长度特征和宽度特征。
  19. 根据权利要求18所述的装置，其特征在于，所述装置还包括：
    合并模块，用于基于所述第一角点特征对应的长度特征和宽度特征以及所述第二角点特征对应的长度特征和宽度特征，确定所述目标检测对象在所述检测图像中的检测框；确定任意两个相互重叠的检测框之间的交并比；在所述交并比大于预设阈值的情况下，将所述任意两个重叠的检测框合并为一个检测框。
  20. 根据权利要求11至19任意一项所述的装置，其特征在于，所述分类模块，具体用于，
    对所述对象特征进行至少一级卷积处理，得到所述目标检测对象属于至少一个预设类别的概率；
    根据所述目标检测对象属于至少一个预设类别的概率，从所述预设类别中确定所述目标检测对象的类别。
  21. 一种电子设备，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为调用所述存储器存储的指令，以执行权利要求1至10中任意一项所述的方法。
  22. 一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现权利要求1至10中任意一项所述的方法。
  23. 一种计算机程序，所述计算机程序包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现权利要求1至10中任意一项所述的方法。
PCT/CN2019/127181 2019-10-29 2019-12-20 目标检测方法及装置、电子设备和存储介质 WO2021082231A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2022519510A JP2022549728A (ja) 2019-10-29 2019-12-20 目標検出方法及び装置、電子機器、並びに記憶媒体
KR1020227009421A KR20220050960A (ko) 2019-10-29 2019-12-20 목표 검출 방법, 장치, 전자 기기 및 기억 매체

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911038042.3A CN110796649B (zh) 2019-10-29 2019-10-29 目标检测方法及装置、电子设备和存储介质
CN201911038042.3 2019-10-29

Publications (1)

Publication Number Publication Date
WO2021082231A1 true WO2021082231A1 (zh) 2021-05-06

Family

ID=69441803

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/127181 WO2021082231A1 (zh) 2019-10-29 2019-12-20 目标检测方法及装置、电子设备和存储介质

Country Status (5)

Country Link
JP (1) JP2022549728A (zh)
KR (1) KR20220050960A (zh)
CN (1) CN110796649B (zh)
TW (1) TWI772757B (zh)
WO (1) WO2021082231A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113591843A (zh) * 2021-07-12 2021-11-02 中国兵器工业计算机应用技术研究所 仿初级视觉皮层的目标检测方法、装置及设备

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111401224B (zh) * 2020-03-13 2023-05-23 北京字节跳动网络技术有限公司 目标检测方法、装置及电子设备
CN111582263A (zh) * 2020-05-12 2020-08-25 上海眼控科技股份有限公司 车牌识别方法、装置、电子设备及存储介质
CN113066086B (zh) * 2021-04-26 2022-08-30 深圳市商汤科技有限公司 道路病害检测方法及装置、电子设备和存储介质
CN113140005B (zh) * 2021-04-29 2024-04-16 上海商汤科技开发有限公司 目标对象定位方法、装置、设备及存储介质
CN113936042B (zh) * 2021-12-16 2022-04-05 深圳佑驾创新科技有限公司 一种目标跟踪方法、装置和计算机可读存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2010182078A (ja) * 2009-02-05 2010-08-19 Olympus Corp 画像処理装置及び画像処理プログラム
CN108648163A (zh) * 2018-05-17 2018-10-12 厦门美图之家科技有限公司 一种人脸图像的增强方法及计算设备
CN108846440A (zh) * 2018-06-20 2018-11-20 腾讯科技（深圳）有限公司 图像处理方法及装置、计算机可读介质和电子设备
CN109034266A (zh) * 2018-08-16 2018-12-18 新智数字科技有限公司 一种目标图像检测方法、装置及***
CN109118456A (zh) * 2018-09-26 2019-01-01 北京字节跳动网络技术有限公司 图像处理方法和装置
CN109344727A (zh) * 2018-09-07 2019-02-15 苏州创旅天下信息技术有限公司 身份证文本信息检测方法及装置、可读存储介质和终端
CN109344762A (zh) * 2018-09-26 2019-02-15 北京字节跳动网络技术有限公司 图像处理方法和装置

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108107571B (zh) * 2013-10-30 2021-06-01 株式会社摩如富 图像处理装置及方法及非暂时性计算机可读记录介质
EP3511903A4 (en) * 2016-09-12 2019-10-02 Panasonic Intellectual Property Management Co., Ltd. DEVICE FOR PRODUCING A THREE-DIMENSIONAL MODEL AND METHOD FOR GENERATING A THREE-DIMENSIONAL MODEL
US10402943B2 (en) * 2016-10-20 2019-09-03 Htc Corporation Image enhancement device and method for convolutional network apparatus
US9946539B1 (en) * 2017-05-23 2018-04-17 Google Llc Accessing data in multi-dimensional tensors using adders
CN109145928B (zh) * 2017-06-16 2020-10-27 杭州海康威视数字技术股份有限公司 一种基于图像的车头朝向识别方法及装置
CN107909093B (zh) * 2017-10-27 2021-02-02 浙江大华技术股份有限公司 一种物品检测的方法及设备
CN108305260B (zh) * 2018-03-02 2022-04-12 苏州大学 一种图像中角点的检测方法、装置及设备
CN108520229B (zh) * 2018-04-04 2020-08-07 北京旷视科技有限公司 图像检测方法、装置、电子设备和计算机可读介质
CN109977997B (zh) * 2019-02-13 2021-02-02 中国科学院自动化研究所 基于卷积神经网络快速鲁棒的图像目标检测与分割方法
CN110136136B (zh) * 2019-05-27 2022-02-08 北京达佳互联信息技术有限公司 场景分割方法、装置、计算机设备及存储介质

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2010182078A (ja) * 2009-02-05 2010-08-19 Olympus Corp 画像処理装置及び画像処理プログラム
CN108648163A (zh) * 2018-05-17 2018-10-12 厦门美图之家科技有限公司 一种人脸图像的增强方法及计算设备
CN108846440A (zh) * 2018-06-20 2018-11-20 腾讯科技（深圳）有限公司 图像处理方法及装置、计算机可读介质和电子设备
CN109034266A (zh) * 2018-08-16 2018-12-18 新智数字科技有限公司 一种目标图像检测方法、装置及***
CN109344727A (zh) * 2018-09-07 2019-02-15 苏州创旅天下信息技术有限公司 身份证文本信息检测方法及装置、可读存储介质和终端
CN109118456A (zh) * 2018-09-26 2019-01-01 北京字节跳动网络技术有限公司 图像处理方法和装置
CN109344762A (zh) * 2018-09-26 2019-02-15 北京字节跳动网络技术有限公司 图像处理方法和装置

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113591843A (zh) * 2021-07-12 2021-11-02 中国兵器工业计算机应用技术研究所 仿初级视觉皮层的目标检测方法、装置及设备
CN113591843B (zh) * 2021-07-12 2024-04-09 中国兵器工业计算机应用技术研究所 仿初级视觉皮层的目标检测方法、装置及设备

Also Published As

Publication number Publication date
CN110796649B (zh) 2022-08-30
CN110796649A (zh) 2020-02-14
JP2022549728A (ja) 2022-11-28
TWI772757B (zh) 2022-08-01
TW202117595A (zh) 2021-05-01
KR20220050960A (ko) 2022-04-25

Similar Documents

Publication Publication Date Title
WO2021082231A1 (zh) 目标检测方法及装置、电子设备和存储介质
CN109255352B (zh) 目标检测方法、装置及***
TWI766201B (zh) 活體檢測方法、裝置以及儲存介質
WO2021022983A1 (zh) 图像处理方法和装置、电子设备、计算机可读存储介质
US8406554B1 (en) Image binarization based on grey membership parameters of pixels
CN111583097A (zh) 图像处理方法、装置、电子设备及计算机可读存储介质
WO2019214320A1 (zh) 车辆损伤识别的处理方法、处理设备、客户端及服务器
CN112784670A (zh) 基于像素差异的对象检测
CN110163188B (zh) 视频处理以及在视频中嵌入目标对象的方法、装置和设备
WO2021035812A1 (zh) 一种图像处理方法及装置、电子设备和存储介质
Yeh et al. Real-time background modeling based on a multi-level texture description
JP2014531097A (ja) マルチレイヤ連結成分をヒストグラムと共に用いるテキスト検出
Xiao et al. Defocus blur detection based on multiscale SVD fusion in gradient domain
WO2022217876A1 (zh) 实例分割方法及装置、电子设备及存储介质
US20220254251A1 (en) Method of recognizing illegal parking of vehicle, device and storage medium
CN114663952A (zh) 对象分类方法、深度学习模型的训练方法、装置和设备
Zhang et al. A crowd counting framework combining with crowd location
WO2019188053A1 (en) Method, system and computer readable medium for integration and automatic switching of crowd estimation techniques
CN110879986A (zh) 人脸识别的方法、设备和计算机可读存储介质
CN117218622A (zh) 路况检测方法、电子设备及存储介质
WO2018058573A1 (zh) 对象检测方法、对象检测装置以及电子设备
WO2022111688A1 (zh) 人脸活体检测方法、装置及存储介质
CN111340044A (zh) 图像处理方法、装置、电子设备及存储介质
CN112052863B (zh) 一种图像检测方法及装置、计算机存储介质、电子设备
CN113256662A (zh) 病理切片图像分割方法、装置、计算机设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19950615

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 20227009421

Country of ref document: KR

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 2022519510

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19950615

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 05.09.2022)

122 Ep: pct application non-entry in european phase

Ref document number: 19950615

Country of ref document: EP

Kind code of ref document: A1