CN116863313B - Target re-identification method and system based on label increment refining and symmetrical scoring - Google Patents
Target re-identification method and system based on label increment refining and symmetrical scoring Download PDFInfo
- Publication number
- CN116863313B CN116863313B CN202311134901.5A CN202311134901A CN116863313B CN 116863313 B CN116863313 B CN 116863313B CN 202311134901 A CN202311134901 A CN 202311134901A CN 116863313 B CN116863313 B CN 116863313B
- Authority
- CN
- China
- Prior art keywords
- sample
- suspicious
- tag
- distance
- samples
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 55
- 238000007670 refining Methods 0.000 title claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims abstract description 62
- 238000009792 diffusion process Methods 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims abstract description 21
- 230000006870 function Effects 0.000 claims description 10
- 238000004891 communication Methods 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 7
- 230000007246 mechanism Effects 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000001174 ascending effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000036039 immunity Effects 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/98—Detection or correction of errors, e.g. by rescanning the pattern or by human intervention; Evaluation of the quality of the acquired patterns
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- 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
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The invention discloses a target re-identification method and a target re-identification system based on label increment refining and symmetrical scoring, wherein the method comprises the following steps: pre-training a first depth model by using data containing noise labels, and extracting the characteristics of all samples through the first depth model; calculating the distance from the extracted sample characteristics to the class center, and dividing the sample into a clean sample and a suspicious sample according to the distance; calculating the characteristic distance from each suspicious sample to a clean sample, and constructing a characteristic distance matrix; generating a tag diffusion weight matrix according to the characteristic distance matrix, and diffusing the tag of the clean sample to the suspicious sample based on the tag diffusion weight matrix to obtain an optimized tag of the suspicious sample; training a second depth model with the clean samples and their labels, the suspicious samples and their optimized labels for target re-identification. According to the method, the sample is divided into the clean sample and the suspicious sample, and the labels of the suspicious sample are refined according to the label scattering weight, so that the label noise is effectively reduced.
Description
Technical Field
The invention belongs to the technical field of target re-identification, and particularly relates to a target re-identification method and system based on label increment refining and symmetrical scoring.
Background
Computer vision is an important branch of the field of artificial intelligence, aimed at using computer algorithms to simulate and understand the human visual system. It obtains useful information by analyzing image and video data to solve various practical problems such as object recognition, object localization, and anomaly detection. Target re-recognition is an important task in the field of computer vision, generally involving recognition of objects such as pedestrians and vehicles. In the task of pedestrian re-identification, the goal is to identify and track the identity of different pedestrians, even though they are present at different times and places, and to be able to track in a video sequence. The vehicle re-identification task is to find a specific vehicle in images shot by different cameras, and further lock and track the same vehicle by searching. The target re-identification task is widely applied to a plurality of fields such as safety monitoring, intelligent transportation and the like, and has important significance for the application of artificial intelligence in modern life.
However, current target re-identification is still highly dependent on large-scale precisely annotated data. Existing large-scale images or videos are usually automatically marked by a computer, but due to factors such as gestures, shielding, deformation, illumination and the like, incorrect marking is easy to occur, and even under the condition of expensive manual marking, similar images are difficult to accurately identify. Therefore, the tag noise problem is difficult to avoid. Conventional target re-identification research is mainly focused on improving identification accuracy or model network, and negative effects caused by tag noise are not fully considered. Some methods with noise immunity in certain fields are often limited to a single task, lacking generalization capability. For example, the tag noise method in the image classification field is not applicable or does not achieve the desired effect in the pedestrian or vehicle re-recognition field. Some approaches in pedestrian or vehicle re-identification tasks also fail to correct false labels or make full use of the relationship information between samples. Recent studies utilize neighborhood label optimization to improve labels, however, when there are more noisy labels, the reliability of the neighborhood labels is not guaranteed.
There are methods for label noise that achieve significant results in image classification tasks, but involve only a few categories in image classification tasks, each category containing thousands of images. However, the task of pedestrian re-identification involves many different identities, with a sample of each identity typically not exceeding 30 images. Thus, a well-behaved method or a tag optimization method in image classification may not be suitable for pedestrian re-recognition tasks with tag noise. Some studies have explored this problem, e.g., purifyNet in PurifyNet: A robust person re-identification model with noisy labels, which proposes a hard-aware instance re-weighting strategy, assigning more weight to difficult samples with correct labels, thus optimizing the loss function; CORE, collaborative refining for person re-identification with label noise, proposes a self-label refinement strategy for joint optimization of labels and networks. These studies all use models to optimize tags, however, the models suffer from tag noise and problems of large intra-class differences, small inter-class differences, etc. due to cross-device presence when predicting the correct tags. The aforementioned methods each treat each sample as an isolated individual, and ignore neighborhood information, which is important in optimizing noise tags. In addition, there are some efforts to target recognition that employ sparse pairwise loss, with a few suitable pairwise adaptive forward mining strategies to dynamically adapt to changes within different categories or to improve the accuracy of target recognition by generating missing image portions. However, these methods do not adequately address the tag noise problems caused by the presence of occlusions, illumination, and cross-device factors.
In view of the foregoing, there is still a need to develop more intensive research and design more applicable methods to solve this problem for the task of re-identifying pedestrians with tag noise.
Disclosure of Invention
In view of the above, the invention provides a target re-identification method and a target re-identification system based on label increment refining and symmetrical scoring, which are used for solving the problem that the influence of label noise cannot be effectively reduced.
The invention discloses a target re-identification method based on label increment refining and symmetrical scoring, which comprises the following steps:
pre-training a first depth model by using data containing noise labels, and extracting the characteristics of all samples through the first depth model;
calculating the distance from the extracted sample characteristics to the class center, and dividing the sample into a clean sample and a suspicious sample according to the distance;
calculating the characteristic distance from each suspicious sample to a clean sample, and constructing a characteristic distance matrix;
generating a tag diffusion weight matrix according to the characteristic distance matrix, and diffusing the tag of the clean sample to the suspicious sample based on the tag diffusion weight matrix to obtain an optimized tag of the suspicious sample;
training a second depth model with the clean samples and their labels, the suspicious samples and their optimized labels for target re-identification.
On the basis of the above technical solution, preferably, calculating the distance from the extracted sample feature to the class center includes:
calculating the distance from the extracted sample feature to the category center;
calculating the average distance from the sample in each category to the center of the category;
and taking the samples with the average distance smaller than or equal to the average distance of the category as clean samples and taking the samples with the average distance larger than the average distance of the category as suspicious samples.
On the basis of the above technical solution, preferably, the calculating the feature distance from each suspicious sample to the clean sample and constructing the feature distance matrix specifically includes:
calculating suspicious samplesiTo clean samplejIs the characteristic distance of (2)：
if k Representative sampleiIs the first of (2)kThe dimensional characteristics of the object are defined,jf k representative samplejIs the first of (2)kThe dimensional characteristics of the object are defined,k=1,2,...,n，nis the dimension of the sample feature;
using a softmax function pairNormalizing to obtain normalized characteristic distance:
constructing an NxN feature distance matrix S from the normalized feature distances, whereinS is the firstiLine 1jColumn elements.
On the basis of the above technical solution, preferably, the formula for generating the tag diffusion weight matrix according to the feature distance matrix is:
wherein the method comprises the steps ofDiffusing weight matrix for tagsPIs the first of (2)iLine 1jColumn elements.
On the basis of the above technical solution, preferably, the diffusing the label of the clean sample to the suspicious sample based on the label diffusing weight matrix specifically includes:
calculating a new tag matrix of suspicious samples:
new tag matrix for suspicious sample, +.>Original tag matrix for suspicious sample, +.>The matrix is N rows and C columns, and C is the total number of sample categories;
for suspicious samplesiOptimized tagThe method comprises the following steps:
wherein,crepresents the firstcThe number of categories of the product,c=1,2,...,C，for new tag matrix->Is the first of (2)iLine 1cColumns.
On the basis of the above technical solution, preferably, in the process of training the second depth model by using the clean sample and the label thereof, the suspicious sample and the optimization label thereof to perform target re-identification, the suspicious sample is weighted, and a weight calculation formula is as follows:
wherein,for suspicious samplesiIndividual weight parameters of->Representing the turn on which the model was trained.
On the basis of the above technical solution, preferably, in the process of training the second depth model by using the clean sample and the label thereof, the suspicious sample and the optimized label thereof to re-identify the target, a loss function is calculated by an individual symmetry scoring mechanism:
wherein,Lrepresenting a loss function and,CE=plnq、RCE=qlnprespectively suspicious samplesiIs a standard cross entropy and an inverse cross entropy of (a),pfor a true tag distribution,qa category distribution is predicted for the model.
In a second aspect of the present invention, a target re-identification system based on tag delta refinement and symmetry scoring is disclosed, the system comprising:
and the feature extraction module is used for: pre-training a first depth model by using data containing noise labels, and extracting the characteristics of all samples through the first depth model;
sample dividing module: the method comprises the steps of calculating the distance from an extracted sample feature to a class center, and dividing the sample into a clean sample and a suspicious sample according to the distance;
and a label optimizing module: the method comprises the steps of calculating the characteristic distance from each suspicious sample to a clean sample, and constructing a characteristic distance matrix; generating a tag diffusion weight matrix according to the characteristic distance matrix, and diffusing the tag of the clean sample to the suspicious sample based on the tag diffusion weight matrix to obtain an optimized tag of the suspicious sample;
target re-identification module: for training a second depth model with the clean samples and their labels, suspicious samples and their optimized labels for target re-identification.
In a third aspect of the present invention, an electronic device is disclosed, comprising: at least one processor, at least one memory, a communication interface, and a bus;
the processor, the memory and the communication interface complete communication with each other through the bus;
the memory stores program instructions executable by the processor which the processor invokes to implement the method according to the first aspect of the invention.
In a fourth aspect of the invention, a computer-readable storage medium is disclosed, storing computer instructions that cause a computer to implement the method according to the first aspect of the invention.
Compared with the prior art, the invention has the following beneficial effects:
1) The method and the device distinguish the clean sample and the suspicious sample by using the extracted characteristics and the intra-class average distance, and refine the label of the suspicious sample according to the similarity of the suspicious sample and the clean sample as the diffusion weight of the clean sample label, thereby reducing the label noise and solving the problem that an optimization algorithm is invalid when the neighborhood labels of the sample are mostly wrong labels when the noise labels are more.
2) According to the invention, the learning degree of the model on the suspicious sample is measured by using the individual symmetrical score in the training process, and the model prediction result is added as the score parameter, so that the learning of the model on the potential noise sample and the hard sample far from the class center is balanced, the fault tolerance rate of the model on the tag noise is improved, the model can be better adapted to the existence of the tag noise, the sensitivity on the wrong tag is reduced as much as possible, and the influence of the potential tag noise on the depth model is further reduced.
Drawings
In order to more clearly illustrate the embodiments of the invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a target re-identification method based on label increment refining and symmetric scoring of the present invention.
Detailed Description
The following description of the embodiments of the present invention will clearly and fully describe the technical aspects of the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present invention without making any inventive effort, are intended to fall within the scope of the present invention.
Referring to fig. 1, the invention provides a target re-identification method based on label increment refining and symmetry scoring, which comprises the following steps:
s1, pre-training a first depth model by using sample data containing noise labels, and extracting the characteristics of all samples through the first depth model.
The resnet50, convolutional neural network, etc. may be employed as a first depth model, trained using sample data containing noise labels, extracting features of all samples.
S2, calculating the distance from the extracted sample characteristic to the class center, and dividing the sample into a clean sample and a suspicious sample according to the distance.
The specific way of dividing into clean samples and suspicious samples is:
calculating Euclidean distance from the extracted sample characteristics to the category center;
calculating the average distance from the sample in each category to the center of the category;
and taking the samples with the average distance smaller than or equal to the average distance of the category as clean samples and taking the samples with the average distance larger than the average distance of the category as suspicious samples.
Because the sample data contains noise labels, the characteristics of the samples are inaccurate, and the accuracy of target re-identification is affected. Therefore, the method and the device distinguish the clean sample from the suspicious sample by calculating the average distance in the class, so that the suspicious sample can be conveniently subjected to label optimization in the follow-up process.
S3, calculating the characteristic distance from each suspicious sample to the clean sample, and constructing a characteristic distance matrix.
For ease of calculation, a matrix S of N x N is provided for preserving the characteristic distance between samples, N being the number of samples,S ij representing suspicious samplesiAnd clean samplejIf the sample isjFor suspicious sampleS ij =0. If a sample isiFor clean samples, only the matrix SS ii =1，iThe other columns of rows are all 0.
Calculating suspicious samples using the following formulaiTo clean samplejIs the characteristic distance of (2)：
Wherein,if k representative sampleiIs the first of (2)kThe dimensional characteristics of the object are defined,jf k representative samplejIs the first of (2)kThe dimensional characteristics of the object are defined,k=1,2,...,n，nis the characteristic dimension of the sample.
Using a softmax function pairNormalizing to obtain normalized characteristic distance:
constructing an NxN feature distance matrix S from the normalized feature distances, whereinS is the firstiLine 1jThe column elements are arranged in a row,i,j=1,2,...,N。
s4, generating a tag diffusion weight matrix according to the feature distance matrix, and diffusing the tags of the clean samples to the suspicious samples based on the tag diffusion weight matrix to obtain optimized tags of the suspicious samples.
The formula for generating the tag diffusion weight matrix according to the characteristic distance matrix is as follows:
wherein the method comprises the steps ofDiffusing weight matrix for tagsPIs the first of (2)iLine 1jColumn element, tag diffusion weight matrixPA matrix of N rows and N columns, whereinP ij Representative sampleiAndjis a similarity of (3). By aligningPEach row is ordered according to the ascending order of the rows and records the front of each row after being orderedkThe number of columns in which each sample is most similar to each sample can be foundkClean samples. If a sample isjNot a sampleiFront of (2)kNeighbors, thenP ij =0. In this way, a neighborhood relationship can be established.
Diffusion of weight matrix through tagsPCalculating a new tag matrix of suspicious samples:
new tag matrix for suspicious sample, +.>Original tag matrix for suspicious sample, +.>The matrix is N rows and C columns, and C is the total number of sample categories;
for suspicious samplesiOptimized tagThe method comprises the following steps:
wherein,crepresents the firstcThe number of categories of the product,c=1,2,...,C，for new tag matrix->Is the first of (2)iLine 1cColumns.
According to the method, the characteristic distances between the clean samples and the suspicious samples are measured to obtain all the clean samples which are most similar to the suspicious samples, the characteristic distances between each suspicious sample and the clean samples are normalized, the similarity obtained according to the normalization result is used as clean sample diffusion label information to generate the weight of a new label for the suspicious samples, and the label type with the largest weight is found according to the label diffusion result to obtain the new optimized label, so that the problem that an optimization algorithm is invalid when the neighborhood labels of the samples are mostly wrong labels when noise labels are more can be solved.
S5, training a second depth model by using the clean sample and the label thereof, the suspicious sample and the optimized label thereof to re-identify the target.
In the process of training a second depth model by using the clean sample and the label thereof, the suspicious sample and the optimized label thereof to carry out target re-identification, the suspicious sample is weighted, and a weight calculation formula is as follows:
wherein,for suspicious samplesiIs a weight parameter of the individual.
Furthermore, for suspicious samples, the loss function is calculated by an individual symmetry scoring mechanism:
wherein,Lrepresenting the loss function, i.e., the individual symmetry score of the measurable sample.CE=plnq、RCE=qlnpRespectively suspicious samplesiIs a standard cross entropy and an inverse cross entropy of (a),representing the turn on which the model is trained,pfor a true tag distribution,qa category distribution is predicted for the model.
For clean samples, the individual symmetry scores default toThe weight defaults to 1.
Under training data containing label noise, the true labels of the model may be erroneous, thus the original distributionpNot necessarily representing the true distribution, in which case the distribution predicted by the modelqInstead, the invention has a certain representativeness, so the invention increases the model prediction result based on the individual symmetry scoring mechanismqAs a means oflnpBy a coefficient of (1)The method realizes the prediction result of the exchange model and the original label distribution to achieve the aim of reducing the harm of label noise, thereby improving the fault tolerance rate of the label noise. In this way, the model can better adapt to the presence of label noise during training and minimize sensitivity to false labels. In addition, the individual symmetry scoring mechanism scores suspicious samples and clean samples respectively, scoring of each individual instance level is achieved, the learning degree of the model is indirectly measured through the scoring, the learning process of difficult samples and samples possibly with error labels is helped to be balanced, more accurate feedback is provided for the model, and the performance and the robustness of the model in a target re-recognition task are improved.
Corresponding to the embodiment of the method, the invention also provides a target re-identification system based on label increment refining and symmetrical scoring, which comprises the following steps:
and the feature extraction module is used for: pre-training a first depth model by using data containing noise labels, and extracting the characteristics of all samples through the first depth model;
sample dividing module: the method comprises the steps of calculating the distance from an extracted sample feature to a class center, and dividing the sample into a clean sample and a suspicious sample according to the distance;
and a label optimizing module: the method comprises the steps of calculating the characteristic distance from each suspicious sample to a clean sample, and constructing a characteristic distance matrix; generating a tag diffusion weight matrix according to the characteristic distance matrix, and diffusing the tag of the clean sample to the suspicious sample based on the tag diffusion weight matrix to obtain an optimized tag of the suspicious sample;
target re-identification module: for training a second depth model with the clean samples and their labels, suspicious samples and their optimized labels for target re-identification.
The system embodiments and the method embodiments are in one-to-one correspondence, and the brief description of the system embodiments is just to refer to the method embodiments.
The invention also discloses an electronic device, comprising: at least one processor, at least one memory, a communication interface, and a bus; the processor, the memory and the communication interface complete communication with each other through the bus; the memory stores program instructions executable by the processor that the processor invokes to implement the aforementioned methods of the present invention.
The invention also discloses a computer readable storage medium storing computer instructions for causing a computer to implement all or part of the steps of the methods of the embodiments of the invention. The storage medium includes: a usb disk, a removable hard disk, a ROM, a RAM, a magnetic disk, or an optical disk, or other various media capable of storing program codes.
The system embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, i.e., may be distributed over a plurality of network elements. One of ordinary skill in the art may select some or all of the modules according to actual needs without performing any inventive effort to achieve the objectives of the present embodiment.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, alternatives, and improvements that fall within the spirit and scope of the invention.
Claims (6)
1. A target re-identification method based on label increment refining and symmetry scoring, the method comprising:
pre-training a first depth model by using data containing noise labels, and extracting the characteristics of all samples through the first depth model;
calculating the distance from the extracted sample characteristics to the class center, and dividing the sample into a clean sample and a suspicious sample according to the distance;
calculating the characteristic distance from each suspicious sample to a clean sample, and constructing a characteristic distance matrix;
generating a tag diffusion weight matrix according to the characteristic distance matrix, and diffusing the tag of the clean sample to the suspicious sample based on the tag diffusion weight matrix to obtain an optimized tag of the suspicious sample;
the formula for generating the tag diffusion weight matrix according to the characteristic distance matrix is as follows:
；
wherein the method comprises the steps ofDiffusing weight matrix for tagsPIs the first of (2)iLine 1jColumn elements; wherein->Is the first of the feature distance matrix SiLine 1jColumn elements;
based on the tag diffusion weight matrix, diffusing the tag of the clean sample to the suspicious sample, and obtaining the optimized tag of the suspicious sample specifically comprises the following steps:
calculating a new tag matrix of suspicious samples:
；
new tag matrix for suspicious sample, +.>Original tag matrix for suspicious sample, +.>The matrix is N rows and C columns, and C is the total number of sample categories;
for suspicious samplesiOptimized tagThe method comprises the following steps:
；
wherein,crepresents the firstcThe number of categories of the product,c=1,2,...,C，for new tag matrix->Is the first of (2)iLine 1cA column;
training a second depth model by using the clean sample and the label thereof, the suspicious sample and the optimized label thereof to carry out target re-identification;
in the process of training the second depth model by using the clean sample and the label thereof, the suspicious sample and the optimized label thereof to re-identify the target, a loss function is calculated through an individual symmetry scoring mechanism:
；
wherein,Lrepresenting a loss function and,CE=plnq、RCE=qlnprespectively suspicious samplesiIs a standard cross entropy and an inverse cross entropy of (a),pfor a true tag distribution,qthe class distribution is predicted for the model,，/>for suspicious samplesiIndividual weight parameters of->Representing the turn on which the model was trained.
2. The target re-identification method based on label increment refining and symmetric scoring according to claim 1, wherein calculating the distance from the extracted sample feature to the class center comprises dividing the sample into a clean sample and a suspicious sample according to the distance size:
calculating the distance from the extracted sample feature to the category center;
calculating the average distance from the sample in each category to the center of the category;
and taking the samples with the average distance smaller than or equal to the average distance of the category as clean samples and taking the samples with the average distance larger than the average distance of the category as suspicious samples.
3. The method for re-identifying a target based on incremental refining and symmetric scoring of claim 1, wherein the calculating the feature distance from each suspicious sample to a clean sample and constructing the feature distance matrix specifically comprises:
calculating suspicious samplesiTo clean samplejIs the characteristic distance of (2)：
；
if k Representative sampleiIs the first of (2)kThe dimensional characteristics of the object are defined,jf k representative samplejIs the first of (2)kThe dimensional characteristics of the object are defined,k=1,2,...,n，nis the dimension of the sample feature;
using a softmax function pairNormalizing to obtain normalized characteristic distance:
；
constructing an NxN feature distance matrix S from the normalized feature distances, whereinS is the firstiLine 1jColumn elements.
4. A target re-identification system based on tag delta refinement and symmetry scoring using the method of any one of claims 1-3, the system comprising:
and the feature extraction module is used for: pre-training a first depth model by using data containing noise labels, and extracting features of all samples through the first depth model;
sample dividing module: the method comprises the steps of calculating the distance from an extracted sample feature to a class center, and dividing the sample into a clean sample and a suspicious sample according to the distance;
and a label optimizing module: the method comprises the steps of calculating the characteristic distance from each suspicious sample to a clean sample, and constructing a characteristic distance matrix; generating a tag diffusion weight matrix according to the characteristic distance matrix, and diffusing the tag of the clean sample to the suspicious sample based on the tag diffusion weight matrix to obtain an optimized tag of the suspicious sample;
target re-identification module: for training a second depth model with the clean samples and their labels, suspicious samples and their optimized labels for target re-identification.
5. An electronic device, comprising: at least one processor, at least one memory, a communication interface, and a bus;
the processor, the memory and the communication interface complete communication with each other through the bus;
the memory stores program instructions executable by the processor, the processor invoking the program instructions to implement the method of any of claims 1-3.
6. A computer readable storage medium storing computer instructions for causing a computer to implement the method of any one of claims 1 to 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311134901.5A CN116863313B (en) | 2023-09-05 | 2023-09-05 | Target re-identification method and system based on label increment refining and symmetrical scoring |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311134901.5A CN116863313B (en) | 2023-09-05 | 2023-09-05 | Target re-identification method and system based on label increment refining and symmetrical scoring |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116863313A CN116863313A (en) | 2023-10-10 |
| CN116863313B true CN116863313B (en) | 2024-01-12 |
Family
ID=88229046
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311134901.5A Active CN116863313B (en) | 2023-09-05 | 2023-09-05 | Target re-identification method and system based on label increment refining and symmetrical scoring |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116863313B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20190096877A (en) * | 2019-07-31 | 2019-08-20 | 엘지전자 주식회사 | Artificial intelligence(ai)-based voice sampling apparatus and method for providing speech style in heterogeneous label |
| CN114842267A (en) * | 2022-05-23 | 2022-08-02 | 南京邮电大学 | Image classification method and system based on label noise domain self-adaption |
| CN115439887A (en) * | 2022-08-26 | 2022-12-06 | 三维通信股份有限公司 | Pedestrian re-identification method and system based on pseudo label optimization and storage medium |
| CN115511012A (en) * | 2022-11-22 | 2022-12-23 | 南京码极客科技有限公司 | Class soft label recognition training method for maximum entropy constraint |
| CN115618236A (en) * | 2022-11-16 | 2023-01-17 | 南京码极客科技有限公司 | Self-supervision distributed internal and external noise identification algorithm |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110826424B (en) * | 2019-10-21 | 2021-07-27 | 华中科技大学 | Pedestrian searching method based on pedestrian re-identification driving positioning adjustment |
-
2023
- 2023-09-05 CN CN202311134901.5A patent/CN116863313B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20190096877A (en) * | 2019-07-31 | 2019-08-20 | 엘지전자 주식회사 | Artificial intelligence(ai)-based voice sampling apparatus and method for providing speech style in heterogeneous label |
| CN114842267A (en) * | 2022-05-23 | 2022-08-02 | 南京邮电大学 | Image classification method and system based on label noise domain self-adaption |
| CN115439887A (en) * | 2022-08-26 | 2022-12-06 | 三维通信股份有限公司 | Pedestrian re-identification method and system based on pseudo label optimization and storage medium |
| CN115618236A (en) * | 2022-11-16 | 2023-01-17 | 南京码极客科技有限公司 | Self-supervision distributed internal and external noise identification algorithm |
| CN115511012A (en) * | 2022-11-22 | 2022-12-23 | 南京码极客科技有限公司 | Class soft label recognition training method for maximum entropy constraint |
Non-Patent Citations (2)
| Title |
|---|
| PurifyNet: A Robust Person Re-Identification Model With Noisy Labels;Mang Ye et al.;《IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY》;第2655-2666页 * |
| 结合伪标签生成与噪声标签学习的弱监督点云分割;邓安等;《计算机辅助设计与图形学学报》;第2节 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116863313A (en) | 2023-10-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Cheng et al. | High-quality proposals for weakly supervised object detection | |
| CN108960140B (en) | Pedestrian re-identification method based on multi-region feature extraction and fusion | |
| CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
| CN109558823B (en) | Vehicle identification method and system for searching images by images | |
| CN111709311B (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| JP2022141931A (en) | Method and device for training living body detection model, method and apparatus for living body detection, electronic apparatus, storage medium, and computer program | |
| CN111881722B (en) | Cross-age face recognition method, system, device and storage medium | |
| CN109190698B (en) | Classification and identification system and method for network digital virtual assets | |
| Liu et al. | Exploiting web images for fine-grained visual recognition by eliminating open-set noise and utilizing hard examples | |
| CN116311214B (en) | License plate recognition method and device | |
| Yao et al. | R²IPoints: Pursuing Rotation-Insensitive Point Representation for Aerial Object Detection | |
| CN114880478A (en) | Weak supervision aspect category detection method based on theme information enhancement | |
| Xiao et al. | An extended attention mechanism for scene text recognition | |
| CN115482418A (en) | Semi-supervised model training method, system and application based on pseudo negative label | |
| Chaitra et al. | Deep-CNNTL: text localization from natural scene images using deep convolution neural network with transfer learning | |
| CN111814653B (en) | Method, device, equipment and storage medium for detecting abnormal behavior in video | |
| CN111898528B (en) | Data processing method, device, computer readable medium and electronic equipment | |
| Liu et al. | A Multi-scale Feature Pyramid SAR Ship Detection Network with Robust Background Interference | |
| CN116863313B (en) | Target re-identification method and system based on label increment refining and symmetrical scoring | |
| CN110287970B (en) | Weak supervision object positioning method based on CAM and covering | |
| CN115640418B (en) | Cross-domain multi-view target website retrieval method and device based on residual semantic consistency | |
| Hu et al. | Count, Decode and Fetch: A New Approach to Handwritten Chinese Character Error Correction | |
| CN113158878B (en) | Heterogeneous migration fault diagnosis method, system and model based on subspace | |
| CN114067159A (en) | EUS-based fine-granularity classification method for submucosal tumors | |
| Deng et al. | Abnormal Occupancy Grid Map Recognition using Attention Network |
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 |