CN110503112A - A kind of small target deteection of Enhanced feature study and recognition methods - Google Patents

A kind of small target deteection of Enhanced feature study and recognition methods Download PDF

Info

Publication number
CN110503112A
CN110503112A CN201910794606.XA CN201910794606A CN110503112A CN 110503112 A CN110503112 A CN 110503112A CN 201910794606 A CN201910794606 A CN 201910794606A CN 110503112 A CN110503112 A CN 110503112A
Authority
CN
China
Prior art keywords
layer
module
small
convolutional layer
small target
Prior art date
Legal status (The legal status 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 status listed.)
Granted
Application number
CN201910794606.XA
Other languages
Chinese (zh)
Other versions
CN110503112B (en
Inventor
程建
林莉
李�灿
周晓晔
李月男
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
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 University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN201910794606.XA priority Critical patent/CN110503112B/en
Publication of CN110503112A publication Critical patent/CN110503112A/en
Application granted granted Critical
Publication of CN110503112B publication Critical patent/CN110503112B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24143—Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of small target deteection of Enhanced feature study and recognition methods, belong to image procossing, pattern-recognition and computer vision field, solve the problems, such as that small target deteection is low with identification mission detection accuracy in the prior art and network efficiency is low.The present invention successively constructs basic network module, characteristic extracting module, candidate frame generation module and prediction output module as small target deteection and identification network；Based on extraction Small object sample image data, and the Small object sample image data of extraction is pre-processed；It obtains being trained in the small target deteection and identification network of pretreated Small object sample image data input initialization parameter, the small target deteection after being trained and identification network；Small object image to be predicted is inputted into trained small target deteection and identification network, by propagated forward, realizes prediction block position and the classification information for exporting Small object end-to-endly.The present invention is for small target deteection and identification.

Description

A kind of small target deteection of Enhanced feature study and recognition methods
Technical field
A kind of small target deteection of Enhanced feature study and recognition methods belong to image for small target deteection and identification Processing, pattern-recognition and computer vision field.
Background technique
So far, object detection and recognition task is still one of popular research direction in computer vision field, by In its relatively broad engineer application, so that the task is rapidly developed and innovates in academic research field.In fact, Object detection and recognition task also has important role in life, such as based on the face under object detection and recognition task Identify the security inspection applications in the important public transport place such as airport, railway station；Based under object detection and recognition task Car plate detection and identification, for specification traffic, detection traffic safety also have practical significance.
Object detection and recognition task with common classification task the difference is that traditional classification task only need it is defeated The probability that single category result out, i.e. the input picture belong to some classification.Therefore, more when existing in a picture to be detected When a interesting target, simple classification task is insufficient for this kind of demand.Opposite, object detection and recognition task can Go out the position of attention object by detecting network positions, and its classification is judged.And for object detection and recognition For subtask --- small target deteection and identification, since neural network is for Small object feature learning scarce capacity, to lead Small target deteection and identification mission is caused all not to be highly improved in recent years.
Traditional object detection and recognition method, is largely all based on anchor mechanism method, that is, in characteristic pattern On design several prediction blocks, prediction block and true frame are compared, using certain pre-designed judgment criteria, chosen A prediction block with the immediate prediction block of true frame as network is selected, and the classification of the prediction block is predicted.With The development of deep learning, object detection and recognition task development and innovation gradually have been obtained in terms of improving performance, at present Detection is broadly divided into two different settling modes with identification mission: 1) the object detection and recognition method of two stages task；2) The object detection and recognition method of Single-phase mission.Specifically, if being two by object detection and recognition Task-decomposing Independent subtask: if detection is handled with classification, this kind of method is referred to as the object detection and recognition of two stages task.Together Reason it is found that if it is be known as based on end-to-end realization object detection and recognition task the target detection of Single-phase mission with Recognition methods.Either single-stage process or dual stage process, all lower for the detection accuracy of Small object object (note: this In Small object object refer to that pixel point areas in the picture is less than the target of 32x32).Reason essentially consists in following two side Face: 1) characteristic present scarce capacity of the feature that neural network learning arrives for Small object object；2) tradition is based on anchor machine The algorithm of target detection of system is using the IOU (i.e. overlapping area size) calculated between prediction block and true frame, then using pre- It is lesser to favored area to reject IOU for the threshold value first set.But for small target deteection, any one characteristic layer it is small Target its to be mapped to the region area in original image generally smaller, using this kind of judgment criteria, small target deteection task is come The problem of saying the detection leakage phenomenon that will appear maximum probability and network efficiency low (i.e. detection speed is slow).
Summary of the invention
Aiming at the problem that the studies above, the purpose of the present invention is to provide a kind of Enhanced feature study small target deteection and Recognition methods, solves that small target deteection is low with identification mission detection accuracy in the prior art and network efficiency is low (detects speed Slowly) the problem of.
In order to achieve the above object, the present invention adopts the following technical scheme:
A kind of small target deteection of Enhanced feature study and recognition methods, include the following steps:
S1, it the basic network module for successively constructing the feature for extracting Small object and exporting preliminary characteristic pattern, is used for Feature is further extracted on the basis of preliminary characteristic pattern and export characteristic extracting module that two hourglass storehouses of characteristic pattern are formed, The candidate frame generation module of candidate frame is generated based on characteristic pattern, the recurrence of prediction block coordinate and prediction block classification are carried out based on candidate frame The prediction output module of classification is as small target deteection and identification network, i.e. deep neural network, and random initializtion is small after building The parameter of object detection and recognition network；
S2, Small object sample image data is extracted based on COCO data set, i.e., extraction pixel point areas is that 32x32 is below Small object sample image data, and the Small object sample image data of extraction is pre-processed；Obtain pretreated Small object It is trained in the small target deteection and identification network of sample image data input initialization parameter, it is small after being trained Object detection and recognition network；
S3, by Small object image input to be predicted trained small target deteection and identification network, by it is preceding to It propagates, realizes prediction block position and the classification information for exporting Small object end-to-endly.
Further, the basic network module in the step S1 is improved ResNet-101 or improved VGG16.
Further, the improved ResNet-101 successively include input layer, first group of convolutional layer, maximum pond layer, Second group of convolutional layer, third group convolutional layer, the 4th group of convolutional layer and the 5th group of convolutional layer, wherein the input of input layer having a size of The image data of 513x513；First group of convolutional layer successively includes 1 7x7 convolution operation and 1 nonlinear activation function operation two Partially, second group of convolutional layer successively includes 9 convolutional layers, a nonlinear activation layer and average pond layer, third group convolutional layer Successively comprising 12 convolutional layers, a nonlinear activation layer and average pond layer, the 4th group of convolutional layer successively includes 69 convolution Layer, a nonlinear activation layer and average pond layer, the 5th group of convolutional layer successively includes 9 convolutional layers, a nonlinear activation Layer and average pond layer, wherein each convolutional layer of second group of convolutional layer into the 5th group of convolutional layer successively passes through 1 1x1 volumes Product, 1 3x3 convolution sum, 1 1x1 convolution operation.
Further, the improved VGG16 successively includes the first convolutional layer, the second convolutional layer, third convolutional layer, the 4th Convolutional layer and the 5th convolutional layer, wherein the first convolutional layer and the second convolutional layer successively include that the convolution that 2 convolution kernels are 3x3 is grasped Make and nonlinear activation function operates, third convolutional layer, Volume Four lamination and the 5th convolutional layer successively include that 3 convolution kernels are The convolution operation and nonlinear activation function of 3x3 operates.
Further, characteristic extracting module and then basic network module in the step S1, the list in characteristic extracting module A hourglass storehouse is made of 3 rank sampling units, is in " hourglass " shape, every rank sampling unit includes that convolution module is reflected with identical Penetrate module；Wherein, convolution module successively includes 1 down-sampling layer, 3 convolutional layers and 1 up-sampling layer, second-order sampling unit In convolution module in second convolutional layer be convolution module in the first rank sampling unit, the volume in third rank sampling unit Second convolutional layer in volume module is the convolution module in second-order sampling unit, convolutional layer that each convolutional layer is 3x3, under adopt Down-sampling ratio in sample layer isWherein, d indicates the d articles branch in convolution module, up-samples and up-samples in layer Using the method for bilinear interpolation；Identical mapping module is used for the input by the down-sampling layer of convolution module and above adopts layer Output carries out jump connection, for learning the detailed information to shallow-layer feature in deep layer network.
Further, the candidate frame generation module is based on anchor generation mechanism, in the feature of characteristic extracting module output Each pixel position of figure can generate 9 various sizes of candidate frames, each candidate frame is mapped on original image i.e. meeting A corresponding candidate frame.
Further, in the step S1, output module and then candidate frame generation module is predicted, wherein carry out prediction block Coordinate recurrence is that prediction block coordinate position is returned out by Smooth L1 loss function, and prediction block category classification refers to based on pre- It surveys characteristic pattern corresponding to frame and corresponding prediction block is obtained by softmax loss function after wx+b returns out a numerical value Classification information, x refers to the pixel point value in characteristic pattern, wherein prediction block refers to candidate frame；It is specific as follows:
Prediction block center position coordinate is x, and y, width and height are respectively w, h, if any one prediction block central point Position and high wide information are xa, ya, wa, ha, true frame center position is xt, yt, wt, ht, prediction block center position are as follows: X, y, w, h, if true frame and candidate frame should offset be g=(gx, gy, gw, gh), specifically solve formula are as follows:
The offset actually obtained is l=(lx, ly, lw, lh), specifically solve formula are as follows:
Prediction block coordinate position is returned out using Smooth L1 loss function, corresponds to solution formula are as follows:
In formula, i indicates all positive sample collection, i.e. any one in prediction block set；
Wherein, Smooth L1 loss function are as follows:
Pass through softmax loss function after wx+b returns out a numerical value based on characteristic pattern corresponding to prediction block The classification information of corresponding prediction block is obtained, softmax loss function is as follows:
Wherein, c is prediction box label, i.e., prediction classification, k are true box label, i.e., Small object sample image data is true Real frame classification, LclsFor Classification Loss function, α is hyper parameter, can be automatically adjusted in experiment, and classification includes plant, television set, ship, Chair.
Further, in the step S1, random initializtion small target deteection and the parameter of identification network refer to using larger Public data to small target deteection and identification network carry out pre-training, obtain the parameter of one group of initialization, wherein biggish public affairs Opening data is lmageNet.
Further, it when carrying out small target deteection and identification network training in the step S2, increases central point and differentiates mould Block block, candidate frame, k near neighbor method and non-maxima suppression method for being generated based on candidate frame generation module are to being predicted The candidate frame that output module is predicted, specific steps are as follows: according to the center point of the true frame of Small object sample image data It sets, k near neighbor method is then utilized around it, candidate frame corresponding to the k center position nearest with it is determined, as first The selected candidate frame of step after the processing of k near neighbor method, then by non-maxima suppression method, obtains optimal candidate frame.
Further, in the step S2, pretreatment is carried out to the Small object sample image data of extraction and is referred to small mesh Mark sample image data carries out positive and negative 90 degree of rotation, random cropping or scaling operation；
The specific implementation process of small target deteection and identification network after being trained are as follows: the Small object pre-processed In the small target deteection and identification network of sample image data input initialization parameter, carries out propagated forward and returned and divide Class as a result, and according to Small object sample image data to return and classification results ask loss, utilize the loss backpropagation update Small target deteection and identification network of network parameter, small target deteection and identification net after reaching iterated conditional, after being trained Network.
The present invention compared with the existing technology, its advantages are shown in:
One, lower for Small object characteristic present scarce capacity and network detection efficiency is that small target deteection and identification are appointed One of more scabrous problem in business, the Small object of the characteristic extracting module proposed by the present invention formed based on two hourglass storehouses Network is detected, it, can be effective with detailed information by semantic information in characteristic extracting module by way of similar pyramid fusion Ground is fused together, to enhance the feature representation ability to Small object, improves low for small target deteection precision ask Topic, wherein also satisfying that meter of the candidate frame of many redundancies to reduce subsequent step to frame is rejected according to central point discrimination module It calculates；Small candidate frame can be prevented to be taken as negative sample to losing in that part NMS.
Two, it present invention introduces being sentenced otherwise according to true frame central point geometric distance, the most may efficiently pick out There is the positive sample frame (prediction block) of target, to reduce redundancy frame for the burden of network query function.
Three, the present invention focuses more on frame around real goal, thus reduce the calculating apart from actual position frame farther out, Therefore it can be improved detection speed.
Detailed description of the invention
Fig. 1 is flow diagram of the invention；
Fig. 2 is the network structure of the feature base module that two hourglass storehouses are formed in the present invention and prediction output module Figure；
Fig. 3 is the schematic diagram of characteristic extracting module in the present invention；
Fig. 4 is the schematic diagram that the central point obtained based on the present invention differentiates prediction block.
Fig. 5 is based on the feature base for using SSD, DSSD, 1 hourglass storehouse to be formed on the basis of basic network of the invention The feature base module effect contrast figure that plinth module and 2 hourglass storehouses are formed；Wherein, One-hourglass indicates only to include 1 The feature base module that a hourglass storehouse is formed；Two-hourglass indicates that proposed by the present invention includes that 2 hourglass storehouses are formed Feature base module, AP indicate mean accuracy, subscript S, M, L respectively indicate small scale, mesoscale and large scale target, SSD Refer to that using a Fusion Features module predicts that shallow-layer and further feature, DSSD refers to using two Fusion Features Shallow-layer feature merge predicting by module by way of deconvolution with further feature.
Fig. 6 is the testing result schematic diagram that the present invention with the prior art SSD and DSSD are shown in the practical example of the present invention, In, (a) indicates SSD network experiment result figure, (b) indicates DSSD network experiment result figure, (c) indicates experimental result of the invention Figure.
Specific embodiment
Below in conjunction with the drawings and the specific embodiments, the invention will be further described.
In order to solve the above-mentioned problem, the present invention forms characteristic extracting module using two storehouse hourglasses, and enhancing is shallow The semantic information of layer feature, shallow-layer feature is merged, enhance the minutia of further feature with further feature at the same time, thus Strengthens network describes the feature of Small object object, improves the precision of small target deteection.On the other hand, improve herein for The judgment criteria of original 1OU is screened and the immediate prediction block of true frame using the evaluation method that central point is concentrated.Last benefit The location information of prediction block and the classification information of prediction block are returned out respectively with prediction discrimination module.
As described in Figure 1, a kind of small target deteection of Enhanced feature study and recognition methods, include the following steps:
S1, it the basic network module for successively constructing the feature for extracting Small object and exporting preliminary characteristic pattern, is used for Feature is further extracted on the basis of preliminary characteristic pattern and export characteristic extracting module that two hourglass storehouses of characteristic pattern are formed, The candidate frame generation module of candidate frame is generated based on characteristic pattern, the recurrence of prediction block coordinate and prediction block classification are carried out based on candidate frame The prediction output module of classification is as small target deteection and identification network, i.e. deep neural network, and random initializtion is small after building The parameter of object detection and recognition network；
Basic network module in the step S1 is improved ResNet-101 or improved VGG16.
As shown in Fig. 2, state improved ResNet-101 successively include input layer, first group of convolutional layer, maximum pond layer, Second group of convolutional layer, third group convolutional layer, the 4th group of convolutional layer and the 5th group of convolutional layer, wherein the input of input layer having a size of The image data of 513x513；First group of convolutional layer successively includes 1 7x7 convolution operation and 1 nonlinear activation function operation two Partially, second group of convolutional layer successively includes 9 convolutional layers, a nonlinear activation layer and average pond layer, third group convolutional layer Successively comprising 12 convolutional layers, a nonlinear activation layer and average pond layer, the 4th group of convolutional layer successively includes 69 convolution Layer, a nonlinear activation layer and average pond layer, the 5th group of convolutional layer successively includes 9 convolutional layers, a nonlinear activation Layer and average pond layer, wherein each convolutional layer of second group of convolutional layer into the 5th group of convolutional layer successively passes through 1 1x1 volumes Product, 1 3x3 convolution sum, 1 1x1 convolution operation.
The improved VGG16 successively includes the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination With the 5th convolutional layer, wherein the first convolutional layer and the second convolutional layer successively include the convolution operation that 2 convolution kernels are 3x3 and Nonlinear activation function operation, third convolutional layer, Volume Four lamination and the 5th convolutional layer successively include that 3 convolution kernels are 3x3 Convolution operation and nonlinear activation function operation.
Single hourglass as shown in figure 3, the characteristic extracting module and then basic network module, in characteristic extracting module Storehouse is made of 3 rank sampling units, is in " hourglass " shape, every rank sampling unit includes convolution module and identical mapping mould Block；Wherein, convolution module successively includes 1 down-sampling layer, 3 convolutional layers and 1 up-sampling layer, in second-order sampling unit Second convolutional layer in convolution module is the convolution module in the first rank sampling unit, the convolution mould in third rank sampling unit Second convolutional layer in block is the convolution module in second-order sampling unit, and each convolutional layer is convolutional layer, the down-sampling layer of 3x3 In down-sampling ratio beWherein, d indicates the d articles branch in convolution module, up-samples and up-samples use in layer Be bilinear interpolation method；Identical mapping module is used to by the input of the down-sampling layer of convolution module and above adopt the output of layer Jump connection is carried out, for learning the detailed information to shallow-layer feature in deep layer network.
The candidate frame generation module is based on anchor generation mechanism, in each of the characteristic pattern of characteristic extracting module output Pixel position can generate 9 various sizes of candidate frames, each candidate frame, which is mapped on original image, can correspond to one Candidate frame.
Predict output module and then candidate frame generation module, wherein carrying out the recurrence of prediction block coordinate is to pass through Smooth L1 loss function returns out prediction block coordinate position, and prediction block category classification refers to be passed through based on characteristic pattern corresponding to prediction block It crosses wx+b and returns out and the classification information of corresponding prediction block is obtained by softmax loss function after a numerical value, x refers to characteristic pattern Pixel point value in the middle, wherein prediction block refers to candidate frame；It is specific as follows:
Prediction block center position coordinate is x, and y, width and height are respectively w, h, if any one prediction block central point Position and high wide information are xa, ya, wa, ha, true frame center position is xt, yt, wt, ht, prediction block center position are as follows: X, y, w, h, if true frame and candidate frame should offset be g=(gx, gy, gw, gh), specifically solve formula are as follows:
The offset actually obtained is l=(lx, ly, lw, lh), specifically solve formula are as follows:
Prediction block coordinate position is returned out using Smooth L1 loss function, corresponds to solution formula are as follows:
In formula, i indicates all positive sample collection, i.e. any one in prediction block set；
Wherein, Smooth L1 loss function are as follows:
Pass through softmax loss function after wx+b returns out a numerical value based on characteristic pattern corresponding to prediction block The classification information of corresponding prediction block is obtained, softmax loss function is as follows:
Wherein, c is prediction label, i.e. prediction classification, and k is true tag, i.e., true classification, α is hyper parameter, can in experiment Automatic adjustment, LclsFor Classification Loss function, classification includes plant, television set, ship, chair etc..
The parameter of random initializtion small target deteection and identification network, which refers to, examines Small object using biggish public data It surveys and carries out pre-training with identification network, obtain the parameter of one group of initialization, wherein biggish public data is lmageNet.
S2, Small object sample image data is extracted based on COCO data set, because accounting for 41% Small object sample in COCO data set This image data, i.e. extraction pixel point areas is 32x32 Small object sample image data below, and to the Small object sample of extraction This image data is pre-processed；Obtain the Small object inspection of pretreated Small object sample image data input initialization parameter It surveys and is trained with identification network, the small target deteection after being trained and identification network；
As shown in figure 4, increasing central point discrimination module block when carrying out small target deteection and identification network training, being used for Based on candidate frame generation module generate candidate frame, k near neighbor method and non-maxima suppression method to obtain prediction output module The candidate frame predicted, specific steps are as follows: according to the center position of the true frame of Small object sample image data, then exist K near neighbor method is utilized around it, determines candidate frame corresponding to the k center position nearest with it, as what is tentatively selected Candidate frame after the processing of k near neighbor method, then by non-maxima suppression method, obtains that optimal candidate frame.Central point Discrimination module can reject calculating of the candidate frame of many redundancies to reduce subsequent step to frame；It can prevent small candidate frame It is taken as negative sample to losing in that part NMS.
To the Small object sample image data of extraction carry out pretreatment refer to Small object sample image data carry out it is positive and negative 90 degree of rotation, random cropping or scaling operation；
The specific implementation process of small target deteection and identification network after being trained are as follows: the Small object pre-processed In the small target deteection and identification network of sample image data input initialization parameter, carries out propagated forward and returned and divide Class as a result, and according to Small object sample image data to return and classification results ask loss, utilize the loss backpropagation update Small target deteection and identification network of network parameter, small target deteection and identification net after reaching iterated conditional, after being trained Network.
S3, by Small object image input to be predicted trained small target deteection and identification network, by it is preceding to It propagates, realizes prediction block position and the classification information for exporting Small object end-to-endly.Trained small target deteection and identification net Network can detect the image of untrue frame, specific as follows:
Test phase:
1) training stage has obtained small target deteection and has identified the weight parameter of network to get trained Small object is arrived Detection and identification network, the Small object image of input prediction；
2) (such as marginal information, color, the shape letter of the essential characteristic by basic network module study to Small object image Breath)；
3) pass through characteristic extracting module, study arrives the profile information of different scale, obtains multiple dimensioned characteristic pattern；
4) after obtaining multiple dimensioned characteristic pattern, for any one characteristic pattern, each pixel position generates 9 Anchor, each anchor, which is mapped on original image, can correspond to candidate frame；
5) coordinate of candidate frame is subjected to recurrence calculating (wx+b) to it using trained weight w and b；
6) fractional value, which will be obtained, by softmax function after returning (namely belongs to the general of any one classification Rate value), make NMS using this probability value；
7) that optimal frame is picked out as final predicted value.
Embodiment
Small object sample image data is extracted as test set, by the Small object image in test set point from COCO data set It is not input in SSD, DSSD and method of the present invention and is detected, obtain Fig. 5 and result shown in fig. 6, wherein Fig. 6 Show three width Small object images in the testing result of SSD, DSSD and method of the present invention.The present invention is either in small mesh Mark, middle target and big target detection on, will the detection accuracy of more prior arts want high, and as can be seen from the figure SSD Network and DSSD network have a large amount of missing inspection for the detection of Small object, and model structure proposed in this paper has Improve to preferable.
The above is only the representative embodiment in the numerous concrete application ranges of the present invention, to protection scope of the present invention not structure At any restrictions.It is all using transformation or equivalence replacement and the technical solution that is formed, all fall within rights protection scope of the present invention it It is interior.

Claims (10)

1. small target deteection and the recognition methods of a kind of Enhanced feature study, which comprises the steps of:
S1, it the basic network module for successively constructing the feature for extracting Small object and exporting preliminary characteristic pattern, is used for preliminary Feature is further extracted on the basis of characteristic pattern and is exported the characteristic extracting module of two hourglass storehouses formation of characteristic pattern, is based on Characteristic pattern generates the candidate frame generation module of candidate frame, carries out the recurrence of prediction block coordinate and prediction block category classification based on candidate frame Prediction output module as small target deteection and identification network, i.e. deep neural network, random initializtion Small object after building The parameter of detection and identification network；
S2, Small object sample image data is extracted based on COCO data set, i.e., extraction pixel point areas is 32x32 small mesh below Sample image data is marked, and the Small object sample image data of extraction is pre-processed；Obtain pretreated Small object sample It is trained in the small target deteection and identification network of image data input initialization parameter, Small object after being trained Detection and identification network；
S3, by Small object image input to be predicted trained small target deteection and identification network, by propagated forward, Realize the prediction block position for exporting Small object end-to-endly and classification information.
2. small target deteection and the recognition methods of a kind of Enhanced feature study according to claim 1, it is characterised in that: institute Stating the basic network module in step S1 is improved ResNet-101 or improved VGG16.
3. small target deteection and the recognition methods of a kind of Enhanced feature study according to claim 2, it is characterised in that: institute Stating improved ResNet-101 successively includes input layer, first group of convolutional layer, maximum pond layer, second group of convolutional layer, third Group convolutional layer, the 4th group of convolutional layer and the 5th group of convolutional layer, wherein image data of the input of input layer having a size of 513x513； First group of convolutional layer successively includes that 1 7x7 convolution operation and 1 nonlinear activation function operate two parts, second group of convolutional layer Successively comprising 9 convolutional layers, a nonlinear activation layer and average pond layer, third group convolutional layer successively includes 12 convolution Layer, a nonlinear activation layer and average pond layer, the 4th group of convolutional layer successively includes 69 convolutional layers, a nonlinear activation Layer and average pond layer, the 5th group of convolutional layer successively include 9 convolutional layers, a nonlinear activation layer and the pond layer that is averaged, In, each convolutional layer of second group of convolutional layer into the 5th group of convolutional layer successively passes through 1 1x1 convolution, 1 3x3 convolution sum 1 1x1 convolution operation.
4. small target deteection and the recognition methods of a kind of Enhanced feature study according to claim 2, it is characterised in that: institute Stating improved VGG16 successively includes the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolution Layer, wherein the first convolutional layer and the second convolutional layer successively include the convolution operation and nonlinear activation that 2 convolution kernels are 3x3 Function operation, third convolutional layer, Volume Four lamination and the 5th convolutional layer successively include 3 convolution kernels be 3x3 convolution operation with And nonlinear activation function operation.
5. small target deteection and the recognition methods of a kind of Enhanced feature study according to any one of claims 1-4, It is characterized in that: characteristic extracting module and then basic network module in the step S1, the single hourglass in characteristic extracting module Storehouse is made of 3 rank sampling units, is in " hourglass " shape, every rank sampling unit includes convolution module and identical mapping mould Block；Wherein, convolution module successively includes 1 down-sampling layer, 3 convolutional layers and 1 up-sampling layer, in second-order sampling unit Second convolutional layer in convolution module is the convolution module in the first rank sampling unit, the convolution mould in third rank sampling unit Second convolutional layer in block is the convolution module in second-order sampling unit, and each convolutional layer is convolutional layer, the down-sampling layer of 3x3 In down-sampling ratio beWherein, d indicates the d articles branch in convolution module, up-samples and up-samples use in layer Be bilinear interpolation method；Identical mapping module is used to by the input of the down-sampling layer of convolution module and above adopt the output of layer Jump connection is carried out, for learning the detailed information to shallow-layer feature in deep layer network.
6. small target deteection and the recognition methods of a kind of Enhanced feature study according to claim 5, it is characterised in that: institute It states candidate frame generation module and is based on anchor generation mechanism, in each pixel position of the characteristic pattern of characteristic extracting module output 9 various sizes of candidate frames can be generated, each candidate frame, which is mapped on original image, can correspond to a candidate frame.
7. small target deteection and the recognition methods of a kind of Enhanced feature study according to claim 6, it is characterised in that: institute It states in step S1, predicts output module and then candidate frame generation module, wherein carrying out the recurrence of prediction block coordinate is to pass through Smooth L1 loss function returns out prediction block coordinate position, and prediction block category classification refers to based on corresponding to prediction block Characteristic pattern obtains the classification information of corresponding prediction block by softmax loss function after wx+b returns out a numerical value, and x is Refer to the pixel point value in characteristic pattern, wherein prediction block refers to candidate frame；It is specific as follows:
Prediction block center position coordinate is x, and y, width and height are respectively w, h, if any one prediction block center position And high wide information is xa, ya, wa, ha, true frame center position is xt, yt, wt, ht, prediction block center position are as follows: and x, y, W, h, if true frame and candidate frame should offset be g=(gx, gy, gw, gh), specifically solve formula are as follows:
The offset actually obtained is l=(lx, ly, lw, lh), specifically solve formula are as follows:
Prediction block coordinate position is returned out using Smooth L1 loss function, corresponds to solution formula are as follows:
In formula, i indicates all positive sample collection, i.e. any one in prediction block set；
Wherein, Smooth L1 loss function are as follows:
It is obtained after wx+b returns out a numerical value by softmax loss function based on characteristic pattern corresponding to prediction block The classification information of corresponding prediction block, softmax loss function are as follows:
Wherein, c is prediction box label, i.e. prediction classification, and k is true box label, the i.e. true frame of Small object sample image data Classification, LclsFor Classification Loss function, α is hyper parameter, can be automatically adjusted in experiment, classification includes plant, television set, ship, chair Son.
8. small target deteection and the recognition methods of a kind of Enhanced feature study according to claim 1, it is characterised in that: institute It states in step S1, the parameter of random initializtion small target deteection and identification network refers to using biggish public data to Small object Detection carries out pre-training with identification network, obtains the parameter of one group of initialization, wherein biggish public data is ImageNet.
9. small target deteection and the recognition methods of a kind of Enhanced feature study according to claim 6, it is characterised in that: institute When stating progress small target deteection in step S2 and identifying network training, central point discrimination module block is increased, for based on candidate Candidate frame, k near neighbor method and the non-maxima suppression method that frame generation module generates are predicted prediction output module is obtained Candidate frame, specific steps are as follows: then sharp around it according to the center position of the true frame of Small object sample image data With k near neighbor method, candidate frame corresponding to the k center position nearest with it is determined, as tentatively selected candidate frame, warp After crossing the processing of k near neighbor method, then by non-maxima suppression method, obtain optimal candidate frame.
10. a kind of small target deteection of Enhanced feature study and recognition methods, feature exist according to claim 1 or described in 9 In: in the step S2, pretreatment is carried out to the Small object sample image data of extraction and is referred to Small object sample image number It is operated according to positive and negative 90 degree of rotation, random cropping or scaling is carried out；
The specific implementation process of small target deteection and identification network after being trained are as follows: the Small object sample pre-processed In the small target deteection and identification network of image data input initialization parameter, carries out propagated forward and returned and classified knot Fruit, and loss is asked to recurrence and classification results according to Small object sample image data, utilize the loss backpropagation to update small mesh Mark detection and identification network of network parameter, small target deteection and identification network after reaching iterated conditional, after being trained.
CN201910794606.XA 2019-08-27 2019-08-27 Small target detection and identification method for enhancing feature learning Active CN110503112B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910794606.XA CN110503112B (en) 2019-08-27 2019-08-27 Small target detection and identification method for enhancing feature learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910794606.XA CN110503112B (en) 2019-08-27 2019-08-27 Small target detection and identification method for enhancing feature learning

Publications (2)

Publication Number Publication Date
CN110503112A true CN110503112A (en) 2019-11-26
CN110503112B CN110503112B (en) 2023-02-03

Family

ID=68589600

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910794606.XA Active CN110503112B (en) 2019-08-27 2019-08-27 Small target detection and identification method for enhancing feature learning

Country Status (1)

Country Link
CN (1) CN110503112B (en)

Cited By (50)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110929795A (en) * 2019-11-28 2020-03-27 桂林电子科技大学 Method for quickly identifying and positioning welding spot of high-speed wire welding machine
CN111080700A (en) * 2019-12-11 2020-04-28 中国科学院自动化研究所 Medical instrument image detection method and device
CN111091091A (en) * 2019-12-16 2020-05-01 北京迈格威科技有限公司 Method, device and equipment for extracting target object re-identification features and storage medium
CN111160372A (en) * 2019-12-30 2020-05-15 沈阳理工大学 Large target identification method based on high-speed convolutional neural network
CN111242026A (en) * 2020-01-13 2020-06-05 中国矿业大学 Remote sensing image target detection method based on spatial hierarchy perception module and metric learning
CN111259763A (en) * 2020-01-13 2020-06-09 华雁智能科技(集团)股份有限公司 Target detection method and device, electronic equipment and readable storage medium
CN111291796A (en) * 2020-01-21 2020-06-16 中国科学技术大学 Sampling-free method used in target detector model training process
CN111310756A (en) * 2020-01-20 2020-06-19 陕西师范大学 Damaged corn particle detection and classification method based on deep learning
CN111444828A (en) * 2020-03-25 2020-07-24 腾讯科技（深圳）有限公司 Model training method, target detection method, device and storage medium
CN111523403A (en) * 2020-04-03 2020-08-11 咪咕文化科技有限公司 Method and device for acquiring target area in picture and computer readable storage medium
CN111524112A (en) * 2020-04-17 2020-08-11 中冶赛迪重庆信息技术有限公司 Steel chasing identification method, system, equipment and medium
CN111563462A (en) * 2020-05-11 2020-08-21 广东博智林机器人有限公司 Image element detection method and device
CN111583204A (en) * 2020-04-27 2020-08-25 天津大学 Organ positioning method of two-dimensional sequence magnetic resonance image based on network model
CN111597945A (en) * 2020-05-11 2020-08-28 济南博观智能科技有限公司 Target detection method, device, equipment and medium
CN111611947A (en) * 2020-05-25 2020-09-01 济南博观智能科技有限公司 License plate detection method, device, equipment and medium
CN111626208A (en) * 2020-05-27 2020-09-04 北京百度网讯科技有限公司 Method and apparatus for detecting small targets
CN111767962A (en) * 2020-07-03 2020-10-13 中国科学院自动化研究所 One-stage target detection method, system and device based on generation countermeasure network
CN111814850A (en) * 2020-06-22 2020-10-23 浙江大华技术股份有限公司 Defect detection model training method, defect detection method and related device
CN111986126A (en) * 2020-07-17 2020-11-24 浙江工业大学 Multi-target detection method based on improved VGG16 network
CN112163530A (en) * 2020-09-30 2021-01-01 江南大学 SSD small target detection method based on feature enhancement and sample selection
CN112215179A (en) * 2020-10-19 2021-01-12 平安国际智慧城市科技股份有限公司 In-vehicle face recognition method, device, apparatus and storage medium
CN112308062A (en) * 2020-11-23 2021-02-02 浙江卡易智慧医疗科技有限公司 Medical image access number identification method in complex background image
CN112364931A (en) * 2020-11-20 2021-02-12 长沙军民先进技术研究有限公司 Low-sample target detection method based on meta-feature and weight adjustment and network model
CN112396126A (en) * 2020-12-02 2021-02-23 中山大学 Target detection method and system based on detection of main stem and local feature optimization
CN112417980A (en) * 2020-10-27 2021-02-26 南京邮电大学 Single-stage underwater biological target detection method based on feature enhancement and refinement
CN112561898A (en) * 2020-12-22 2021-03-26 电子科技大学中山学院 Optical fiber sensor light spot analysis method based on convolutional neural network
CN112634174A (en) * 2020-12-31 2021-04-09 上海明略人工智能(集团)有限公司 Image representation learning method and system
CN112699914A (en) * 2020-12-02 2021-04-23 中山大学 Target detection method and system based on heterogeneous composite backbone
CN112819793A (en) * 2021-02-01 2021-05-18 宁波港信息通信有限公司 Container damage identification method, device, equipment and readable access medium
CN112926383A (en) * 2021-01-08 2021-06-08 浙江大学 Automatic target identification system based on underwater laser image
CN112990263A (en) * 2021-02-08 2021-06-18 武汉工程大学 Method for detecting dense small targets of high-resolution image
CN113033672A (en) * 2021-03-29 2021-06-25 西安电子科技大学 Multi-class optical image rotating target self-adaptive detection method based on feature enhancement
CN113128476A (en) * 2021-05-17 2021-07-16 广西师范大学 Low-power consumption real-time helmet detection method based on computer vision target detection
CN113128308A (en) * 2020-01-10 2021-07-16 中南大学 Pedestrian detection method, device, equipment and medium in port scene
CN113159215A (en) * 2021-05-10 2021-07-23 河南理工大学 Small target detection and identification method based on fast Rcnn
CN113158757A (en) * 2021-02-08 2021-07-23 海信视像科技股份有限公司 Display device and gesture control method
CN113221947A (en) * 2021-04-04 2021-08-06 青岛日日顺乐信云科技有限公司 Industrial quality inspection method and system based on image recognition technology
CN113221731A (en) * 2021-05-10 2021-08-06 西安电子科技大学 Multi-scale remote sensing image target detection method and system
CN113361437A (en) * 2021-06-16 2021-09-07 吉林建筑大学 Method and system for detecting category and position of minimally invasive surgical instrument
CN113361322A (en) * 2021-04-23 2021-09-07 山东大学 Power line target detection method, device and storage medium based on weighted deconvolution layer number improved DSSD algorithm
WO2021239156A1 (en) * 2020-05-29 2021-12-02 中移(上海)信息通信科技有限公司 Traffic target recognition model training method, and traffic target positioning method and apparatus
CN113822277A (en) * 2021-11-19 2021-12-21 万商云集(成都)科技股份有限公司 Illegal advertisement picture detection method and system based on deep learning target detection
CN113850256A (en) * 2021-09-10 2021-12-28 北京理工大学 Target detection and identification method based on FSAF and fast-slow weight
CN113920322A (en) * 2021-10-21 2022-01-11 广东工业大学 Modular robot kinematic chain configuration identification method and system
CN114240844A (en) * 2021-11-23 2022-03-25 电子科技大学 Unsupervised key point positioning and target detection method in medical image
CN114241232A (en) * 2021-11-23 2022-03-25 电子科技大学 Multi-task learning-based camera position identification and body surface anatomical landmark detection method
CN114359742A (en) * 2022-03-21 2022-04-15 济南大学 Weighted loss function calculation method for optimizing small target detection
CN114529951A (en) * 2022-02-22 2022-05-24 北京工业大学 On-site fingerprint feature point extraction method based on deep learning
CN115984846A (en) * 2023-02-06 2023-04-18 山东省人工智能研究院 Intelligent identification method for small target in high-resolution image based on deep learning
CN117994251A (en) * 2024-04-03 2024-05-07 华中科技大学同济医学院附属同济医院 Method and system for evaluating severity of diabetic foot ulcer based on artificial intelligence

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000036524A1 (en) * 1998-12-16 2000-06-22 Sarnoff Corporation Method and apparatus for training a neural network to detect objects in an image
CN108805203A (en) * 2018-06-11 2018-11-13 腾讯科技（深圳）有限公司 Image procossing and object recognition methods, device, equipment and storage medium again
CN108960212A (en) * 2018-08-13 2018-12-07 电子科技大学 Based on the detection of human joint points end to end and classification method
CN109117876A (en) * 2018-07-26 2019-01-01 成都快眼科技有限公司 A kind of dense small target deteection model building method, model and detection method
CN109344821A (en) * 2018-08-30 2019-02-15 西安电子科技大学 Small target detecting method based on Fusion Features and deep learning
CN109784476A (en) * 2019-01-12 2019-05-21 福州大学 A method of improving DSOD network
CN109977812A (en) * 2019-03-12 2019-07-05 南京邮电大学 A kind of Vehicular video object detection method based on deep learning

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000036524A1 (en) * 1998-12-16 2000-06-22 Sarnoff Corporation Method and apparatus for training a neural network to detect objects in an image
CN108805203A (en) * 2018-06-11 2018-11-13 腾讯科技（深圳）有限公司 Image procossing and object recognition methods, device, equipment and storage medium again
CN109117876A (en) * 2018-07-26 2019-01-01 成都快眼科技有限公司 A kind of dense small target deteection model building method, model and detection method
CN108960212A (en) * 2018-08-13 2018-12-07 电子科技大学 Based on the detection of human joint points end to end and classification method
CN109344821A (en) * 2018-08-30 2019-02-15 西安电子科技大学 Small target detecting method based on Fusion Features and deep learning
CN109784476A (en) * 2019-01-12 2019-05-21 福州大学 A method of improving DSOD network
CN109977812A (en) * 2019-03-12 2019-07-05 南京邮电大学 A kind of Vehicular video object detection method based on deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
H. CLARK-YOUNGER ET AL.: "Stacked Hourglass CNN for Handwritten Character Location", 《2018 INTERNATIONAL CONFERENCE ON IMAGE AND VISION COMPUTING NEW ZEALAND (IVCNZ)》 *
郭之先: "基于深度卷积神经网络的小目标检测", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑(月刊)，2018年第08期》 *

Cited By (78)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110929795A (en) * 2019-11-28 2020-03-27 桂林电子科技大学 Method for quickly identifying and positioning welding spot of high-speed wire welding machine
CN110929795B (en) * 2019-11-28 2022-09-13 桂林电子科技大学 Method for quickly identifying and positioning welding spot of high-speed wire welding machine
CN111080700A (en) * 2019-12-11 2020-04-28 中国科学院自动化研究所 Medical instrument image detection method and device
CN111091091A (en) * 2019-12-16 2020-05-01 北京迈格威科技有限公司 Method, device and equipment for extracting target object re-identification features and storage medium
CN111160372A (en) * 2019-12-30 2020-05-15 沈阳理工大学 Large target identification method based on high-speed convolutional neural network
CN111160372B (en) * 2019-12-30 2023-04-18 沈阳理工大学 Large target identification method based on high-speed convolutional neural network
CN113128308B (en) * 2020-01-10 2022-05-20 中南大学 Pedestrian detection method, device, equipment and medium in port scene
CN113128308A (en) * 2020-01-10 2021-07-16 中南大学 Pedestrian detection method, device, equipment and medium in port scene
CN111242026B (en) * 2020-01-13 2022-07-12 中国矿业大学 Remote sensing image target detection method based on spatial hierarchy perception module and metric learning
CN111259763A (en) * 2020-01-13 2020-06-09 华雁智能科技(集团)股份有限公司 Target detection method and device, electronic equipment and readable storage medium
CN111242026A (en) * 2020-01-13 2020-06-05 中国矿业大学 Remote sensing image target detection method based on spatial hierarchy perception module and metric learning
CN111259763B (en) * 2020-01-13 2024-02-02 华雁智能科技(集团)股份有限公司 Target detection method, target detection device, electronic equipment and readable storage medium
CN111310756A (en) * 2020-01-20 2020-06-19 陕西师范大学 Damaged corn particle detection and classification method based on deep learning
CN111291796A (en) * 2020-01-21 2020-06-16 中国科学技术大学 Sampling-free method used in target detector model training process
CN111444828B (en) * 2020-03-25 2023-06-20 腾讯科技（深圳）有限公司 Model training method, target detection method, device and storage medium
CN111444828A (en) * 2020-03-25 2020-07-24 腾讯科技（深圳）有限公司 Model training method, target detection method, device and storage medium
CN111523403B (en) * 2020-04-03 2023-10-20 咪咕文化科技有限公司 Method and device for acquiring target area in picture and computer readable storage medium
CN111523403A (en) * 2020-04-03 2020-08-11 咪咕文化科技有限公司 Method and device for acquiring target area in picture and computer readable storage medium
CN111524112A (en) * 2020-04-17 2020-08-11 中冶赛迪重庆信息技术有限公司 Steel chasing identification method, system, equipment and medium
CN111524112B (en) * 2020-04-17 2023-04-07 中冶赛迪信息技术(重庆)有限公司 Steel chasing identification method, system, equipment and medium
CN111583204A (en) * 2020-04-27 2020-08-25 天津大学 Organ positioning method of two-dimensional sequence magnetic resonance image based on network model
CN111583204B (en) * 2020-04-27 2022-10-14 天津大学 Organ positioning method of two-dimensional sequence magnetic resonance image based on network model
CN111597945B (en) * 2020-05-11 2023-08-18 济南博观智能科技有限公司 Target detection method, device, equipment and medium
CN111597945A (en) * 2020-05-11 2020-08-28 济南博观智能科技有限公司 Target detection method, device, equipment and medium
CN111563462A (en) * 2020-05-11 2020-08-21 广东博智林机器人有限公司 Image element detection method and device
CN111611947B (en) * 2020-05-25 2024-04-09 济南博观智能科技有限公司 License plate detection method, device, equipment and medium
CN111611947A (en) * 2020-05-25 2020-09-01 济南博观智能科技有限公司 License plate detection method, device, equipment and medium
CN111626208A (en) * 2020-05-27 2020-09-04 北京百度网讯科技有限公司 Method and apparatus for detecting small targets
WO2021239156A1 (en) * 2020-05-29 2021-12-02 中移(上海)信息通信科技有限公司 Traffic target recognition model training method, and traffic target positioning method and apparatus
CN113743163A (en) * 2020-05-29 2021-12-03 中移(上海)信息通信科技有限公司 Traffic target recognition model training method, traffic target positioning method and device
CN111814850A (en) * 2020-06-22 2020-10-23 浙江大华技术股份有限公司 Defect detection model training method, defect detection method and related device
CN111767962B (en) * 2020-07-03 2022-11-08 中国科学院自动化研究所 One-stage target detection method, system and device based on generation countermeasure network
CN111767962A (en) * 2020-07-03 2020-10-13 中国科学院自动化研究所 One-stage target detection method, system and device based on generation countermeasure network
CN111986126B (en) * 2020-07-17 2022-05-24 浙江工业大学 Multi-target detection method based on improved VGG16 network
CN111986126A (en) * 2020-07-17 2020-11-24 浙江工业大学 Multi-target detection method based on improved VGG16 network
CN112163530A (en) * 2020-09-30 2021-01-01 江南大学 SSD small target detection method based on feature enhancement and sample selection
CN112163530B (en) * 2020-09-30 2024-04-09 江南大学 SSD small target detection method based on feature enhancement and sample selection
CN112215179B (en) * 2020-10-19 2024-04-19 平安国际智慧城市科技股份有限公司 In-vehicle face recognition method, device, apparatus and storage medium
CN112215179A (en) * 2020-10-19 2021-01-12 平安国际智慧城市科技股份有限公司 In-vehicle face recognition method, device, apparatus and storage medium
CN112417980A (en) * 2020-10-27 2021-02-26 南京邮电大学 Single-stage underwater biological target detection method based on feature enhancement and refinement
CN112364931B (en) * 2020-11-20 2024-03-19 长沙军民先进技术研究有限公司 Few-sample target detection method and network system based on meta-feature and weight adjustment
CN112364931A (en) * 2020-11-20 2021-02-12 长沙军民先进技术研究有限公司 Low-sample target detection method based on meta-feature and weight adjustment and network model
CN112308062B (en) * 2020-11-23 2022-08-23 浙江卡易智慧医疗科技有限公司 Medical image access number identification method in complex background image
CN112308062A (en) * 2020-11-23 2021-02-02 浙江卡易智慧医疗科技有限公司 Medical image access number identification method in complex background image
CN112396126B (en) * 2020-12-02 2023-09-22 中山大学 Target detection method and system based on detection trunk and local feature optimization
CN112699914B (en) * 2020-12-02 2023-09-22 中山大学 Target detection method and system based on heterogeneous composite trunk
CN112396126A (en) * 2020-12-02 2021-02-23 中山大学 Target detection method and system based on detection of main stem and local feature optimization
CN112699914A (en) * 2020-12-02 2021-04-23 中山大学 Target detection method and system based on heterogeneous composite backbone
CN112561898A (en) * 2020-12-22 2021-03-26 电子科技大学中山学院 Optical fiber sensor light spot analysis method based on convolutional neural network
CN112634174B (en) * 2020-12-31 2023-12-12 上海明略人工智能(集团)有限公司 Image representation learning method and system
CN112634174A (en) * 2020-12-31 2021-04-09 上海明略人工智能(集团)有限公司 Image representation learning method and system
CN112926383B (en) * 2021-01-08 2023-03-03 浙江大学 Automatic target identification system based on underwater laser image
CN112926383A (en) * 2021-01-08 2021-06-08 浙江大学 Automatic target identification system based on underwater laser image
CN112819793A (en) * 2021-02-01 2021-05-18 宁波港信息通信有限公司 Container damage identification method, device, equipment and readable access medium
CN112990263A (en) * 2021-02-08 2021-06-18 武汉工程大学 Method for detecting dense small targets of high-resolution image
CN113158757A (en) * 2021-02-08 2021-07-23 海信视像科技股份有限公司 Display device and gesture control method
CN113033672A (en) * 2021-03-29 2021-06-25 西安电子科技大学 Multi-class optical image rotating target self-adaptive detection method based on feature enhancement
CN113033672B (en) * 2021-03-29 2023-07-28 西安电子科技大学 Multi-class optical image rotation target self-adaptive detection method based on feature enhancement
CN113221947A (en) * 2021-04-04 2021-08-06 青岛日日顺乐信云科技有限公司 Industrial quality inspection method and system based on image recognition technology
CN113361322A (en) * 2021-04-23 2021-09-07 山东大学 Power line target detection method, device and storage medium based on weighted deconvolution layer number improved DSSD algorithm
CN113221731A (en) * 2021-05-10 2021-08-06 西安电子科技大学 Multi-scale remote sensing image target detection method and system
CN113159215A (en) * 2021-05-10 2021-07-23 河南理工大学 Small target detection and identification method based on fast Rcnn
CN113221731B (en) * 2021-05-10 2023-10-27 西安电子科技大学 Multi-scale remote sensing image target detection method and system
CN113128476A (en) * 2021-05-17 2021-07-16 广西师范大学 Low-power consumption real-time helmet detection method based on computer vision target detection
CN113361437A (en) * 2021-06-16 2021-09-07 吉林建筑大学 Method and system for detecting category and position of minimally invasive surgical instrument
CN113850256A (en) * 2021-09-10 2021-12-28 北京理工大学 Target detection and identification method based on FSAF and fast-slow weight
CN113920322A (en) * 2021-10-21 2022-01-11 广东工业大学 Modular robot kinematic chain configuration identification method and system
CN113822277B (en) * 2021-11-19 2022-02-18 万商云集(成都)科技股份有限公司 Illegal advertisement picture detection method and system based on deep learning target detection
CN113822277A (en) * 2021-11-19 2021-12-21 万商云集(成都)科技股份有限公司 Illegal advertisement picture detection method and system based on deep learning target detection
CN114240844A (en) * 2021-11-23 2022-03-25 电子科技大学 Unsupervised key point positioning and target detection method in medical image
CN114241232A (en) * 2021-11-23 2022-03-25 电子科技大学 Multi-task learning-based camera position identification and body surface anatomical landmark detection method
CN114529951B (en) * 2022-02-22 2024-04-02 北京工业大学 On-site fingerprint feature point extraction method based on deep learning
CN114529951A (en) * 2022-02-22 2022-05-24 北京工业大学 On-site fingerprint feature point extraction method based on deep learning
CN114359742A (en) * 2022-03-21 2022-04-15 济南大学 Weighted loss function calculation method for optimizing small target detection
CN114359742B (en) * 2022-03-21 2022-09-16 济南大学 Weighted loss function calculation method for optimizing small target detection
CN115984846B (en) * 2023-02-06 2023-10-10 山东省人工智能研究院 Intelligent recognition method for small targets in high-resolution image based on deep learning
CN115984846A (en) * 2023-02-06 2023-04-18 山东省人工智能研究院 Intelligent identification method for small target in high-resolution image based on deep learning
CN117994251A (en) * 2024-04-03 2024-05-07 华中科技大学同济医学院附属同济医院 Method and system for evaluating severity of diabetic foot ulcer based on artificial intelligence

Also Published As

Publication number Publication date
CN110503112B (en) 2023-02-03

Similar Documents

Publication Publication Date Title
CN110503112A (en) A kind of small target deteection of Enhanced feature study and recognition methods
CN109766830A (en) A kind of ship seakeeping system and method based on artificial intelligence image procossing
CN106127204B (en) A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks
CN105894045B (en) A kind of model recognizing method of the depth network model based on spatial pyramid pond
CN109815886A (en) A kind of pedestrian and vehicle checking method and system based on improvement YOLOv3
CN109543606A (en) A kind of face identification method that attention mechanism is added
CN109977918A (en) A kind of target detection and localization optimization method adapted to based on unsupervised domain
CN108427920A (en) A kind of land and sea border defense object detection method based on deep learning
CN107229904A (en) A kind of object detection and recognition method based on deep learning
Luo et al. Learning high-level features by fusing multi-view representation of MLS point clouds for 3D object recognition in road environments
CN107885764A (en) Based on the quick Hash vehicle retrieval method of multitask deep learning
CN110245655A (en) A kind of single phase object detecting method based on lightweight image pyramid network
CN107609525A (en) Remote Sensing Target detection method based on Pruning strategy structure convolutional neural networks
CN107818302A (en) Non-rigid multiple dimensioned object detecting method based on convolutional neural networks
CN106354735A (en) Image target searching method and device
Huang et al. Immature apple detection method based on improved Yolov3
CN111611998A (en) Adaptive feature block extraction method based on candidate region area and width and height
CN110991257B (en) Polarized SAR oil spill detection method based on feature fusion and SVM
CN110348437A (en) It is a kind of based on Weakly supervised study with block the object detection method of perception
CN110211127B (en) Image partition method based on bicoherence network
CN113379771A (en) Hierarchical human body analytic semantic segmentation method with edge constraint
CN110084284A (en) Target detection and secondary classification algorithm and device based on region convolutional neural networks
CN107392254A (en) A kind of semantic segmentation method by combining the embedded structural map picture from pixel
CN110443155A (en) A kind of visual aid identification and classification method based on convolutional neural networks
CN109711431A (en) The method for tracking target of local block convolution, system and storage medium at one

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant