CN113673589A - Label selection self-adaptive increment detection method and system based on frame distance measurement - Google Patents
Label selection self-adaptive increment detection method and system based on frame distance measurement Download PDFInfo
- Publication number
- CN113673589A CN113673589A CN202110926238.7A CN202110926238A CN113673589A CN 113673589 A CN113673589 A CN 113673589A CN 202110926238 A CN202110926238 A CN 202110926238A CN 113673589 A CN113673589 A CN 113673589A
- Authority
- CN
- China
- Prior art keywords
- old
- target
- label
- frame
- detection
- 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.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 79
- 238000005259 measurement Methods 0.000 title claims abstract description 13
- 230000003044 adaptive effect Effects 0.000 claims abstract description 25
- 238000000034 method Methods 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims description 28
- 238000004422 calculation algorithm Methods 0.000 claims description 18
- 230000007246 mechanism Effects 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 12
- 238000004590 computer program Methods 0.000 claims description 11
- 238000000605 extraction Methods 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000000717 retained effect Effects 0.000 claims description 2
- 238000012216 screening Methods 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 9
- 238000002372 labelling Methods 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 238000012937 correction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000005764 inhibitory process Effects 0.000 description 1
- 238000003780 insertion Methods 0.000 description 1
- 230000037431 insertion Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 206010027175 memory impairment Diseases 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
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
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a label selection adaptive increment detection method and system based on frame distance measurement. The method can detect targets with different scales and is suitable for incremental target detection without old sample data.
Description
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a label selection adaptive increment detection method and system based on frame distance measurement.
Background
The target detection supports wide application as a visual task, and comprises the fields of automatic driving, medical images, man-machine interaction, high-speed rail contact network foreign matter detection and the like, and has irreplaceable effects on a plurality of application programs. In recent years, with the development of deep learning, the target detection technology has made remarkable progress, but a complex system requires a large amount of training time to learn the model, and the final detector performance depends on the availability of a representative set of training examples to a large extent. In practical applications, it is time consuming and expensive to obtain representative data, and thus it is not uncommon for data to appear in small batches over time.
Over time, new classes of targets may appear for which the model has never been learned before, and if the model is directly fine-tuned with new data, catastrophic forgetfulness occurs, and the detection performance of the old class severely degrades while also affecting the detection accuracy of the new class. If a method of combining all data to learn a training model from the beginning is adopted, not only is time-consuming and expensive to calculate, but also an old data set may not be obtained due to privacy, storage and the like.
In recent years, learners use models to screen out a batch of representative old samples to be stored, and train the models together with new data to adapt to changes of the data, but with the increase of the data, the learners need very large memory overhead, and cannot meet actual demands. It is therefore desirable to design an incremental detection method that can better detect objects without having to access old data.
Disclosure of Invention
The invention aims to provide a label selection self-adaptive increment detection method and system based on frame distance measurement.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
a label selection adaptive increment detection method based on frame distance measurement comprises the following steps:
acquiring an old type detection model, acquiring prior information of an old type by using an old model detection data set, and generating a pseudo label for training a detection network by combining a label selection algorithm and a data set label;
according to the generated pseudo label, adopting jump connection attention to train a feature extractor for keeping the important features of the old category and self-adapting all parameters of the training network;
and detecting the preset multi-scale anchor frame by using the trained detection network to obtain the category and position information of the target.
A tag selection adaptive incremental detection system based on a bounding box distance metric, comprising:
a tag selection module: obtaining information of old class targets in the data set by using an old model, and obtaining pseudo labels for training an incremental detection model by using a label selection algorithm based on frame distance measurement;
the feature extraction module based on the attention mechanism comprises: important characteristic information of the old category is learned and kept through a jump attention module;
a detection network module: with the trained detection model, for a batch of input new class images, the class and position information of all targets including the new class and the old class can be detected.
A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the above-mentioned tag selection adaptive increment detection method based on a bezel distance metric when executing the computer program.
A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the above-mentioned tag selection adaptive incremental detection method based on a bounding box distance metric.
Compared with the prior art, the invention has the following remarkable advantages: (1) the invention provides a frame distance measurement-based label selection adaptive increment detection method, which adopts a pseudo label obtained by a label selection algorithm, adopts a multi-scale candidate frame, adopts a jump connection attention mechanism to extract characteristics, learns and retains important characteristics of an old category, adopts a multi-scale network adaptive training new category, and can detect targets with different scales. (2) The trained network can realize incremental detection of all targets, including category and location information, without accessing old data.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a schematic diagram of the system of the present invention.
FIG. 3 is a flow chart of a tag selection algorithm of the present invention.
Detailed Description
A label selection adaptive increment detection method based on frame distance measurement comprises the following steps,
acquiring an old category detection model; acquiring prior information of an old category by using an old model M detection data set; generating a pseudo label by combining a label selection algorithm with a data set label for training a detection network;
according to the generated pseudo label G, adopting jump connection attention to train a feature extractor for keeping the important features of the old category;
the features learned by the old model M are unchanged in the first stage to obtain the important features of the old class, and the attention module used in the second stage retains the key information of the old class and adaptively trains all parameters of the network.
And detecting the preset multi-scale anchor frame by using the trained detection network to obtain the category and position information of the target.
Further, the process of obtaining the old category prior information is as follows:
through a trained old model, setting a proper threshold value and then detecting a data set to obtain the prior information of the old class target, including the class and the position information.
Further, the tag selection process is as follows:
for the obtained information of the old type of target, only frame coordinate information is adopted in the process, the real label of the new type of target in the data set is read at the same time, the result obtained through calculation of a label selection algorithm is compared with a set threshold value, then frame information exceeding the threshold value is removed until all labels are completely selected, and the last old type of label and the real label of the data set are combined to be used as a pseudo label G of the training incremental detection model.
G=YUY″
Y' is the prior information of the old class target processed by the label selection algorithm, namely the category and the frame coordinate.
The label selection algorithm flow is shown in fig. 3, where X represents a data set sample, the data set X only includes label information of a new category, Y is a real label of a new target in the data set, M is an old category detection model, Y ' is label information of an old category obtained through the model M, Y and Y ' are respectively a real frame coordinate of the new category target in the new data sample and an old category target frame coordinate obtained through the old model, h is a preset threshold, and labels satisfying conditions are removed according to a result calculated by a Generalized Intersection (GIOU) method on a union of metric frame distances and a set threshold, until all labels to be selected are completely selected, and finally Y ' and the real label Y of the data set are combined to obtain a pseudo label G for training the incremental detection model.
Further, a jump attention mechanism is used to learn the important features of the old category, wherein the jump attention mechanism is,
attention Module Attention is the Attention processing of the channel Domain first on the input feature F to obtain Mch(F) Then, carrying out element-wise multiplication with the characteristic F; during multiplication, the channel attention value propagates along the spatial dimension; for such modules, insertion is performed on three levels in the feature extraction module respectively, and a final jump attention mechanism is formed and used for learning and retaining the key feature information of the old category.
And (3) carrying out adaptive learning training on the important features learned by the attention module to obtain a detection network, wherein a loss function of the detection network is as follows:
Lloss=Llocal+Lconf+Lprob
Llocal=λ∑∑(2-w×h)(1-CIoU)
Lconf=-∑∑Iobj[C′log(C)+(1-C′)log(1-C)]-λ∑∑Inoobj[C′log(C)+(1-C′)log(1-C)]
Lprob=∑Iobj∑[p′(c)log(p(c))+(1-p′(c))log(1-p(c))]
wherein λ represents the confidence, w and h are the width and height of the picture, the accuracy of the target position is calculated by adopting a frame distance calculation method CIoU focusing on the aspect ratio, C and C 'represent the confidence of the target and the result of network prediction respectively, and p' represent the class probability of the target and the result of network prediction respectively. I isobjThe value is 1 or 0, namely whether a target exists in the cell or not, the value of the target in the cell is 1, otherwise, the value is 0; i isnoobj: there is no target in the cell, which has a value of 1, otherwise it is 0.
The invention also provides an adaptive increment detection system based on label selection, which comprises,
a tag selection module: obtaining information of old class targets in the data set by using an old model, and obtaining pseudo labels for training by using a label selection algorithm;
the feature extraction module based on the attention mechanism comprises: important characteristic information of the old category is learned and kept through a jump attention module;
a detection network module: with the trained detection model, for a batch of input new class images, the class and position information of all targets including the new class and the old class can be detected.
The tag selection module includes:
a detection module: detecting the type and frame information of the old type target according to a preset multi-scale anchor frame;
a screening module: and calculating a frame distance value according to the acquired frame coordinates and the real tags of the data set, removing the frame coordinates exceeding a threshold value, and finally combining the frame coordinates with the real tags of the data set.
After the feature information of the target is obtained, the important features of the old category are kept through the jump attention, and the key feature information of the old category can be obtained in the later model training process, so that the network can learn and detect the target in a self-adaptive manner.
The loss function of the detection network is:
Lloss=Llocal+Lconf+Lprob
wherein L islocalIs the target frame loss, LconfIs the target confidence loss, LprobIs the target class loss and the loss function of the detection network is the sum of the three losses.
The trained network can realize incremental detection of all targets, including category and location information, without accessing old data.
A computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform a bezel distance metric based tag selection adaptive delta detection method.
A computing device comprising one or more processors, memory, and one or more programs stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing a bezel distance metric based tag selection adaptive delta detection method.
The invention is further described below with reference to the accompanying drawings. The following embodiments are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
Example 1
As shown in fig. 2, a tag selection adaptive increment detection method based on a frame distance metric includes the following steps:
step 1: and acquiring prior knowledge of the old category by using the old model detection data set, and then acquiring a pseudo label for training the model by combining with a real label of the new data through a label selection algorithm.
Acquiring a new data set as a training sample, as shown in fig. 1, acquiring labeling information of an old target by using an old model, including category and position information, including the following steps:
11) and predicting the data set by the old type detection model to obtain a detection result (the type and frame coordinates of the old target) and storing the detection result as a label of the old type.
12) Acquiring real labels of a data set, combining labels of old categories, and generating pseudo labels for training through a label selection algorithm, wherein the pseudo labels comprise the following steps:
A1) reading a real label (including category information and position information) of a data set, and presetting a GIOU threshold;
A2) the information of the old type of targets is obtained, the process only adopts frame coordinate information, simultaneously the reality of the new type of targets in the data set is read, the result obtained by calculating through a label selection algorithm is compared with a set threshold value, and then the frame information exceeding the threshold value is removed until all labels are selected;
A3) combining the remaining labeling information of all the old categories with the real labels of the data set to be used as the pseudo labels G of the training model, and calculating as follows:
G=YUY″
wherein Y is the real label of the data set, and Y' is the prior information of the old class target processed by the label selection algorithm, namely the category and the frame coordinate.
Step 2: obtaining a pseudo label according to a label selection algorithm, and then performing feature selection; the method comprises the following steps:
21) and generating a multi-scale pre-selection frame of the feature map by using preset multi-scale according to the size of the feature map, wherein the multi-scale pre-selection is performed by taking a result obtained by clustering the data set as a preset value and adopting a multi-scale anchor frame to adapt to targets with different sizes.
And (2) randomly zooming to any one of preset scale sizes for the same input image according to the pseudo label for training obtained in the step (1), wherein different scales of preselected frames exist for different feature graphs, and each grid unit in the feature graph corresponds to 3 preselected frames and corresponds to 3 scales.
22) And calculating the IOU value of the multi-scale pre-selection frame and the real label of the feature map, reserving the candidate frame exceeding a threshold (generally set to be 0.5), and reserving the candidate frame with the maximum threshold if the condition is not met. By adopting multi-scale region extraction, targets with different scales can be accurately detected, meanwhile, candidate frames are not required to be obtained through later learning, and the target detection speed is increased.
23) Initializing a new model by using the weight of an old model, extracting features by adopting a multi-scale and jump attention mechanism, and learning important features of the old class in the process of extracting the features; the method comprises the following specific steps:
B1) the weights of the old model are loaded, excluding the last layer of the output layer. In the first stage of training the network, parameters of a feature extraction layer are not trained, the feature extraction weight of an old model is used, and meanwhile, an attention module is led to learn important features of old categories;
B2) the attention module firstly processes the attention of the channel domain on the input feature F, and obtainsThen carrying out element-wise multiplication with the feature F to obtain the learning of the attention of the feature channel, and then transmitting the channel attention value along the spatial dimension to obtain the learning of the attention of the feature channel
B3) Such attention modules are inserted at three levels in the feature extraction module, respectively, and the learning of attention is propagated along three scales, and the key feature information of the old category is learned and retained by such a jump attention mechanism.
And step 3: learning the characteristics of the target by using a characteristic extraction module and a jump attention machine, then transmitting the characteristics on a network with three scales, and finally outputting a prediction result with three scales to obtain the category information and the position information of the target; the method comprises the following steps:
31) acquiring a candidate frame of a data set, transmitting the weight acquired by the feature selection module on three layers of networks, training a learning detection target, acquiring position offset data on three scales, and finally acquiring a final predicted frame through correction and IOU calculation; the method comprises the following specific steps:
C1) the learned feature extraction weights are propagated forwards, the weights are trained on three levels, meanwhile, the weights on a single level are transmitted to the next level, and the two levels of features are connected by the level and are used for obtaining feature information of a shallow level and semantic information of a deep level;
C2) outputting a prediction result on three scales through pyramid-type multi-level learning, and correcting the result through a function to obtain a prediction category and a frame coordinate;
32) obtaining the category information and the position information of network prediction, comparing the result with the real label of the data set, and obtaining the frame loss, the confidence coefficient loss and the category loss of model detection as follows:
Llocal=λ∑∑(2-w×h)(1-CIoU)
Lconf=-∑∑Iobj[C′log(C)+(1-C′)log(1-C)]-λ∑∑Inoobj[C′log(C)+(1-C′)log(1-C)]
Lprob=∑Iobj∑[p′(c)log(p(c))+(1-p′(c))log(1-p(c))]
wherein, λ represents the confidence, w and h are the width and height of the picture, the accuracy of the target position is calculated by using CIoU, C and C 'represent the confidence of the target and the result of network prediction respectively, and p' represent the class probability of the target and the result of network prediction respectively.
33) When the loss function result tends to converge, the training is finished, the obtained result is used as the weight of the incremental detection network, and the category and the position information of the target are obtained according to the detection network. When the target is detected, because redundant frames predicted by the network exist, the frames with high overlapping degree are removed by adopting non-maximum value inhibition, and the best position detection result is obtained.
The method adopts the pseudo label obtained by the label selection algorithm, adopts a multi-scale candidate box, adopts a jump connection attention mechanism to extract the characteristics, learns and retains the important characteristics of the old category, adopts a multi-scale network self-adaptive new training category to detect the targets with different scales, and is suitable for incremental target detection without the old sample data.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The present invention is not limited to the above embodiments, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are included in the scope of the claims of the present invention which are filed as the application.
Claims (10)
1. A label selection adaptive increment detection method based on frame distance measurement is characterized by comprising the following steps:
acquiring an old type detection model, acquiring prior information of an old type by using an old model detection data set, and generating a pseudo label for training a detection network by combining a label selection algorithm and a data set label;
according to the generated pseudo label, adopting jump connection attention to train a feature extractor for keeping the important features of the old category and self-adapting all parameters of the training network;
and detecting the preset multi-scale anchor frame by using the trained detection network to obtain the category and position information of the target.
2. The label selection adaptive increment detection method based on the frame distance measurement as claimed in claim 1, wherein the process of obtaining the old category prior information is as follows:
setting a threshold value through a trained old model, and then detecting a data set to obtain prior information of old targets, including category and position information.
3. The border distance metric-based label selection adaptive increment detection method according to claim 1, wherein the label selection process is as follows:
and for the obtained information of the old type of target, only adopting frame coordinate information, simultaneously reading the real label of the new type of target in the data set, comparing the result obtained by calculating through a label selection algorithm with a set threshold value, then removing the frame information exceeding the threshold value until all labels are completely selected, and combining the last old type of label with the real label of the data set to be used as a pseudo label of the training incremental detection model.
4. The tag selection adaptive increment detection method based on the frame distance metric, according to claim 1, is characterized in that a jump attention mechanism is adopted to learn important features of old categories, wherein the jump attention mechanism is as follows:
attention Module Attention is the Attention processing of the channel Domain first on the input feature F to obtain Mch(F) Then, carrying out element-wise multiplication with the characteristic F; during multiplication, the channel attention value propagates along the spatial dimension; and respectively inserting the three layers in the feature extraction module to form a final jump attention mechanism for learning and retaining the key feature information of the old category.
5. The label selection adaptive increment detection method based on the frame distance metric according to claim 1, characterized in that a detection network is obtained by using the important feature adaptive learning training learned by the attention module, and the loss function of the detection network is:
Lloss=Llocal+Lconf+Lprob
Llocal=λ∑∑(2-w×h)(1-CIoU)
Lconf=-∑∑Iobj[C′log(C)+(1-C′)log(1-C)]-λ∑∑Inoobj[C′log(C)+(1-C′)log(1-C)]
Lprob=∑Iobj∑[p′(c)log(p(c))+(1-p′(c))log(1-p(c))]
wherein L islocalIs the target frame loss, LconfIs the target confidence loss, LprobThe method comprises the steps of calculating the accuracy of a target position by adopting a frame distance calculation method CIoU focusing on the length-width ratio, wherein the method comprises the steps of determining the loss of a target class, expressing the confidence coefficient by lambda, calculating the width and the height of a picture by w and h, calculating the accuracy of the target position by C and C 'respectively expressing the confidence coefficient of the target and the result of network prediction, and expressing the class probability of the target and the result of network prediction by p and p' respectively.
6. A tag selection adaptive incremental detection system based on a bounding box distance metric, comprising:
a tag selection module: obtaining information of old class targets in the data set by using an old model, and obtaining pseudo labels for training an incremental detection model by using a label selection algorithm based on frame distance measurement;
the feature extraction module based on the attention mechanism comprises: important characteristic information of the old category is learned and kept through a jump attention module;
a detection network module: and detecting the category and position information of all targets including the new category and the old category for the input new category image by using the trained detection model.
7. The bezel-distance-metric-based tag selection adaptive incremental detection system of claim 6, wherein the tag selection module comprises:
a detection module: detecting the type and frame information of the old type target according to a preset multi-scale anchor frame;
a screening module: and calculating the frame distance according to the acquired frame coordinates and the real tags of the data set, removing the frame coordinates exceeding the threshold value, and finally combining the frame coordinates with the real tags of the data set.
8. The label selection adaptive increment detection system based on the frame distance measurement as claimed in claim 6, wherein after the feature information of the target is obtained, the important features of the old category are retained through the jump attention, and the key feature information of the old category can be obtained in the process of training the model later, so that the network adaptively learns to detect the target;
the loss function of the detection network is:
Lloss=Llocal+Lconf+Lprob
wherein L islocalIs the target frame loss, LconfIs the target confidence loss, LprobIs the target class loss and the loss function of the detection network is the sum of the three losses.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the computer program implements the steps of the bezel distance metric based tag selection adaptive incremental detection method of any of claims 1-5.
10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the bezel distance metric based tag selection adaptive incremental detection method as claimed in any one of claims 1 to 5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110926238.7A CN113673589A (en) | 2021-08-12 | 2021-08-12 | Label selection self-adaptive increment detection method and system based on frame distance measurement |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110926238.7A CN113673589A (en) | 2021-08-12 | 2021-08-12 | Label selection self-adaptive increment detection method and system based on frame distance measurement |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113673589A true CN113673589A (en) | 2021-11-19 |
Family
ID=78542555
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110926238.7A Pending CN113673589A (en) | 2021-08-12 | 2021-08-12 | Label selection self-adaptive increment detection method and system based on frame distance measurement |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113673589A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115438755A (en) * | 2022-11-08 | 2022-12-06 | 腾讯科技(深圳)有限公司 | Incremental training method and device of classification model and computer equipment |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108764269A (en) * | 2018-04-03 | 2018-11-06 | 华南理工大学 | A kind of cross datasets pedestrian recognition methods again based on space-time restriction incremental learning |
CN109614968A (en) * | 2018-10-10 | 2019-04-12 | 浙江大学 | A kind of car plate detection scene picture generation method based on multiple dimensioned mixed image stylization |
CN110059672A (en) * | 2019-04-30 | 2019-07-26 | 福州大学 | A method of microscopic cell image detection model is carried out using incremental learning to increase class study |
CN110837856A (en) * | 2019-10-31 | 2020-02-25 | 深圳市商汤科技有限公司 | Neural network training and target detection method, device, equipment and storage medium |
CN111667016A (en) * | 2020-06-12 | 2020-09-15 | 中国电子科技集团公司第三十六研究所 | Incremental information classification method based on prototype |
CN111967325A (en) * | 2020-07-16 | 2020-11-20 | 长沙理工大学 | Unsupervised cross-domain pedestrian re-identification method based on incremental optimization |
CN112329827A (en) * | 2020-10-26 | 2021-02-05 | 同济大学 | Increment small sample target detection method based on meta-learning |
CN112819110A (en) * | 2021-04-19 | 2021-05-18 | 中国科学院自动化研究所 | Incremental small sample target detection method and system based on weight generation |
-
2021
- 2021-08-12 CN CN202110926238.7A patent/CN113673589A/en active Pending
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108764269A (en) * | 2018-04-03 | 2018-11-06 | 华南理工大学 | A kind of cross datasets pedestrian recognition methods again based on space-time restriction incremental learning |
CN109614968A (en) * | 2018-10-10 | 2019-04-12 | 浙江大学 | A kind of car plate detection scene picture generation method based on multiple dimensioned mixed image stylization |
CN110059672A (en) * | 2019-04-30 | 2019-07-26 | 福州大学 | A method of microscopic cell image detection model is carried out using incremental learning to increase class study |
CN110837856A (en) * | 2019-10-31 | 2020-02-25 | 深圳市商汤科技有限公司 | Neural network training and target detection method, device, equipment and storage medium |
CN111667016A (en) * | 2020-06-12 | 2020-09-15 | 中国电子科技集团公司第三十六研究所 | Incremental information classification method based on prototype |
CN111967325A (en) * | 2020-07-16 | 2020-11-20 | 长沙理工大学 | Unsupervised cross-domain pedestrian re-identification method based on incremental optimization |
CN112329827A (en) * | 2020-10-26 | 2021-02-05 | 同济大学 | Increment small sample target detection method based on meta-learning |
CN112819110A (en) * | 2021-04-19 | 2021-05-18 | 中国科学院自动化研究所 | Incremental small sample target detection method and system based on weight generation |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115438755A (en) * | 2022-11-08 | 2022-12-06 | 腾讯科技(深圳)有限公司 | Incremental training method and device of classification model and computer equipment |
CN115438755B (en) * | 2022-11-08 | 2024-04-02 | 腾讯科技(深圳)有限公司 | Incremental training method and device for classification model and computer equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109117848B (en) | Text line character recognition method, device, medium and electronic equipment | |
CN109117831B (en) | Training method and device of object detection network | |
US20180211393A1 (en) | Image guided video semantic object segmentation method and apparatus | |
CN110363220B (en) | Behavior class detection method and device, electronic equipment and computer readable medium | |
CN110059637B (en) | Face alignment detection method and device | |
CN111488873B (en) | Character level scene text detection method and device based on weak supervision learning | |
CN111681204B (en) | CT rib fracture focus relation modeling method and device based on graph neural network | |
CN111461121A (en) | Electric meter number identification method based on YO L OV3 network | |
CN113569852A (en) | Training method and device of semantic segmentation model, electronic equipment and storage medium | |
CN111368634A (en) | Human head detection method, system and storage medium based on neural network | |
CN114429577B (en) | Flag detection method, system and equipment based on high confidence labeling strategy | |
CN113673589A (en) | Label selection self-adaptive increment detection method and system based on frame distance measurement | |
CN114596440A (en) | Semantic segmentation model generation method and device, electronic equipment and storage medium | |
CN116805387B (en) | Model training method, quality inspection method and related equipment based on knowledge distillation | |
CN112434585A (en) | Method, system, electronic device and storage medium for identifying virtual reality of lane line | |
CN111832435A (en) | Beauty prediction method and device based on migration and weak supervision and storage medium | |
CN114241411B (en) | Counting model processing method and device based on target detection and computer equipment | |
CN116091784A (en) | Target tracking method, device and storage medium | |
CN116977249A (en) | Defect detection method, model training method and device | |
CN116168201B (en) | Lane line segmentation method and device without accurate data labeling | |
CN112131418A (en) | Target labeling method, target labeling device and computer-readable storage medium | |
CN117333730B (en) | Product surface abnormality detection method and terminal | |
CN116543162B (en) | Image segmentation method and system based on feature difference and context awareness consistency | |
CN115731588B (en) | Model processing method and device | |
CN116258924A (en) | Weak supervision learning method and device based on duplex similarity measurement |
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 |