CN103793721B - Pedestrian repeat recognition method and system based on area related feedback - Google Patents
Pedestrian repeat recognition method and system based on area related feedback Download PDFInfo
- Publication number
- CN103793721B CN103793721B CN201410076028.3A CN201410076028A CN103793721B CN 103793721 B CN103793721 B CN 103793721B CN 201410076028 A CN201410076028 A CN 201410076028A CN 103793721 B CN103793721 B CN 103793721B
- Authority
- CN
- China
- Prior art keywords
- region
- feedback
- weight
- query
- similarity
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 62
- 238000005259 measurement Methods 0.000 claims abstract description 19
- 230000014509 gene expression Effects 0.000 claims abstract description 15
- 239000013598 vector Substances 0.000 claims description 36
- 238000004364 calculation method Methods 0.000 claims description 24
- 238000010801 machine learning Methods 0.000 claims description 18
- 238000000691 measurement method Methods 0.000 claims description 16
- 238000012163 sequencing technique Methods 0.000 claims description 13
- 230000000007 visual effect Effects 0.000 claims description 7
- 238000011524 similarity measure Methods 0.000 claims description 5
- 230000001419 dependent effect Effects 0.000 claims 6
- 230000006870 function Effects 0.000 claims 1
- 238000005516 engineering process Methods 0.000 abstract description 2
- 230000003993 interaction Effects 0.000 description 3
- 230000002452 interceptive effect Effects 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000011840 criminal investigation Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011112 process operation Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000001568 sexual effect Effects 0.000 description 1
- 238000012549 training Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a pedestrian repeat recognition method and system based on area related feedback. Initial query matching and feedback sample collection are carried out, irrelevant images are selected as feedback samples, and types of the feedback samples are marked; adjacent sets are confirmed, and area weight and feature weight are adjusted; feature expression and distance measurement are carried out to obtain query matching results; the matching results are inquired, the results are output if requirements are met, and return is carried out to update the feedback samples in an iteration mode if the requirements are not met until the requirements are met. The pedestrian repeat recognition method and system are based on the area related feedback technology, local feature information of pedestrian images is fully utilized, the local features are combined with other information to dynamically adjusting the weight of the local features in real time, and the operation of accurately and quickly finding and successfully matching a target criminal suspect is finally achieved by combining a traditional pedestrian repeat recognition method.
Description
Technical Field
The invention relates to a process for re-identifying a target suspect in a surveillance video in the field of video investigation, and belongs to a pedestrian re-identification method and system based on region-related feedback.
Background
With the wide construction of safe cities and the popularization of monitoring in various places, the quantity of video monitoring data becomes larger and larger, which brings great challenges to criminal investigation and case solving, and the key to solve the case is to quickly and accurately extract target suspects from the mass databases.
The traditional pedestrian re-identification method can effectively solve the problems of missed detection and false detection possibly caused by long-time manual retrieval, but has relatively low matching efficiency, is mainly used for improving the original retrieval sequence by a method of feature expression and distance measurement, and is a non-interactive query method. In recent years, some interactive correlation feedback methods are applied to a pedestrian re-identification system, but most of the methods are based on overall correlation matching of positive sample images, and influence of irrelevant samples in a pedestrian sample library and local features thereof on improvement of query sequencing results is not considered.
In the existing pedestrian re-identification method, the method based on feature expression is widely applied in practice. In the process of pedestrian re-identification, a proper expression mode is found out by extracting different appearance characteristics and motion information characteristics of the target, and then samples with similar or same characteristic expressions are directly matched in a pedestrian sample library until a suspected target is found out.
Patent number "CN 102663366A", entitled "pedestrian object recognition method and system", proposes a pedestrian object recognition method based on feature expression, which includes the steps of collecting video frames, extracting HOG features of the video frames, etc., the extracted video frames include LBP features of pedestrian direction and intensity information, and then recognizing specific pedestrian objects in the surveillance video scene according to the HOG features and the LBP features, the algorithm of the method is simple, the efficiency is high, but the robustness is not good, the method is sensitive to the shooting orientation angle of the camera and different illumination condition changes, which easily causes mismatching, and is not suitable for pedestrian re-recognition in complex environment;
methods of relevant feedback are also widely used in content-based image retrieval. Different from the specific application range of pedestrian re-identification, in the content-based image retrieval, a large number of positive sample images for matching exist, so that a large number of irrelevant sample images are easy to learn, in the human-computer interaction process, corresponding feature expressions are selected according to the information content and the image scene contained in relevant features, and are fed back to the identification subsystem after comparison, matching and sequencing, so that the purpose of optimizing the initial retrieval result is realized. The method has good realization effect and robustness, but needs to train and learn all characteristics of a large number of samples, has high system overhead, relatively complex algorithm and difficult meeting of requirements on real-time property, and is not suitable for being applied to a pedestrian re-identification system.
Patent number "CN 101539930 a", entitled "a relevant feedback image retrieval method", realizes retrieval of a target image by an image retrieval method based on segment similarity measurement and toronto joint feedback, and the matching effect retrieved by this method is good, but the operation process is relatively complex, and it needs to repeatedly train and learn a plurality of features in the image, and is not suitable for actual combat deployment in a pedestrian re-recognition system.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a pedestrian re-identification method and system based on region-related feedback.
The technical scheme of the invention provides a pedestrian re-identification method based on region-related feedback, which comprises the following steps,
step S1, performing initial query matching and feedback sample collection, including the following sub-steps,
step S1.1, performing primary query matching, including taking an input target person image as a query image, performing initial query and outputting an initial query sorting result;
step S1.2, feedback sample collection is carried out, and when the step S1.2 is executed for the first time, irrelevant images are selected from a certain preset number of images which are ranked the most front from the primary sequencing result to serve as feedback samples and mark types to form a feedback sample set; when step S1.2 is executed subsequently, an irrelevant image is selected as a feedback sample from the query ranking result obtained by the previous iteration of step S4, and the type is marked, and a feedback sample set is added;
the marking type mode is that the method is divided into U areas 1 and 2 … U, each feedback sample is marked as one of 2U types, namely similar to the query image based on the area 1, dissimilar based on the area 1, similar based on the area 2, … dissimilar based on the area U, and dissimilar based on the area U; when a feedback sample is marked, extracting visual features in each region according to region division, setting to obtain an M-dimensional feature vector, wherein any dimension is marked as an M-th dimension, and respectively performing similarity comparison with a corresponding region in a query image according to the feature vector;
step S2, determining neighbor set, adjusting region weight and adjusting feature weight, includes the following sub-steps,
s2.1, for the query image, firstly, a sample set based on region similarity and region dissimilarity of region similarity is found out by a region K neighbor set method; then, applying a dynamic k neighbor rule, updating, adjusting and obtaining a new set which contains k neighbors and is similar to the area of each sample marked as a similar area, and updating, adjusting and obtaining a new set which contains k neighbors and is dissimilar to the area of each sample marked as dissimilar to the area;
step S2.2, updating the region weight and the feature weight,
let S be the similarity between the query image p and the image of the ith feedback sample in the set of feedback samplesa(p,gi) The following calculation formula is adopted,
wherein,representing the characteristic weight of the jth region part under the m-dimensional characteristic vector;representing the feature vector of the jth region position in the mth dimension; wj(p,gi) Representing a jth region-based region weight between the query image p and the feedback sample; therein is provided withRepresenting the similarity between the query image p and the feedback sample based on the jth region part under the mth dimension characteristic value; j takes the values of 1 and 2 … U,
at step S2.2.1, using the distance measurement method in machine learning, the region weight is updated by the following calculation formula,
Wj(p,gi)＝Wj(p,gi)×β1β1>1 formula II
Wj(p,gi)＝Wj(p,gi)×β20<β2<1 formula III
Wherein, Wj(p,gi) Area weight representing jth area position, β1、β2Is a preset coefficient;
at step S2.2.2, using the distance measurement method in machine learning, the updated feature weight uses the following calculation formula,
wherein, mum、σmMean values respectively representing characteristic values of the m-th dimensionAnd variance, α is a preset parameter;
step S3, according to the adjusted region weight and the formula I obtained in the step S2, feature expression and distance measurement are carried out, and a query matching result is obtained;
step S4, displaying the query matching result obtained in the step S3, and outputting the result if the query matching result meets the requirement; if not, the step S1.2 is returned to iterate until the requirement is met.
Moreover, when the feedback sample is divided into regions according to the regions and the similarity comparison is respectively carried out on the feedback sample and the corresponding regions in the query image, the similarity measurement adopted by each region is calculated as the following formula,
wherein,representing the characteristic weight of the jth region part, adopting a preset initial value when the step S1.2 is executed for the first time, and executing the weight value updated in the step S2 when the previous iteration is adopted when the step S1.2 is executed subsequently;an m-dimension feature vector representing a j-th region position;representing a query image p and feedback samples giBased on the similarity of the jth region position under the characteristic value of the mth dimension.
Moreover, when 2 regions are divided, including the trunk and the legs, and similarity comparison is performed with the corresponding regions in the query image, the similarity measurement calculation adopted by the trunk and the legs respectively is as follows,
wherein,andrepresenting the characteristic weights of the leg and torso parts respectively,andrespectively representing the m-dimensional characteristic vectors of the leg part and the trunk part;andrespectively representing the characteristic value of the m-dimension between the query image and the feedback sample based on the legDegree of similarity of lower body and m-dimension characteristic value based on body part(ii) similarity of; st(p,gi) Is the value of the torso-part similarity measure, Sl(p,gi) Is the value of the leg portion similarity measure.
Furthermore, the similarity between the query image p and the image of the ith sample in the feedback sample set is calculated by the following formula,
wherein,andfeature weights under the m-dimensional feature vectors representing the leg and torso parts, respectively;andrespectively representing the m-dimensional characteristic vectors of the leg part and the trunk part; wl(p,gi) And Wt(p,gi) Respectively representing leg region and trunk region based weights between the suspect target and the sample;andrespectively representing the characteristic values of the m-dimension between the suspected target and the sample based on the leg regionsDegree of similarity of lower body and m-dimension characteristic value based on body partThe following similarity.
Further, in step S2.2.1, using the distance measure in machine learning, the region weight is updated by using the following calculation formula,
Wt,l(p,gi)＝Wt,l(p,gi)×β1β1>1 formula nine
Wt,l(p,gi)＝Wt,l(p,gi)×β20<β2<Formula 1 ten
Wherein, Wt,l(p,gi) Represents Wt(p,gi) Or Wl(p,gi)，β1、β2Is a preset coefficient.
Furthermore, in step S2.2.2, using the distance measure in machine learning, the updated feature weights are calculated using the following formula,
wherein, mum、σmRespectively representing the mean and variance of the characteristic value of the mth dimension, α is a preset parameter.
The invention also correspondingly provides a pedestrian re-identification system based on the region-related feedback, which comprises the following modules,
the feedback module is used for carrying out primary query matching and feedback sample collection and comprises the following sub-modules,
the primary query matching sub-module is used for taking the input target person image as a query image, performing initial query and outputting an initial query sequencing result;
the feedback sample collection submodule is used for selecting irrelevant images from a certain preset number of images with the top ranking from the primary ranking result as feedback samples and marking the types to form a feedback sample set when the feedback sample collection is executed for the first time; when the feedback sample collection is subsequently executed, irrelevant images are selected from the query sequencing results obtained by the result display module in the previous iteration as feedback samples, the types of the irrelevant images are marked, and the irrelevant images are added into a feedback sample set;
the marking type mode is that the method is divided into U areas 1 and 2 … U, each feedback sample is marked as one of 2U types, namely similar to the query image based on the area 1, dissimilar based on the area 1, similar based on the area 2, … dissimilar based on the area U, and dissimilar based on the area U; when a feedback sample is marked, extracting visual features in each region according to region division, setting to obtain an M-dimensional feature vector, wherein any dimension is marked as an M-th dimension, and respectively performing similarity comparison with a corresponding region in a query image according to the feature vector;
the weight module is used for determining a neighbor set, adjusting the regional weight and adjusting the characteristic weight, and comprises a sub-module for determining the neighbor set, a sub-module for searching a sample set based on regional similarity and regional dissimilarity by a regional K neighbor set method for a query image; then, applying a dynamic k neighbor rule, updating, adjusting and obtaining a new set which contains k neighbors and is similar to the area of each sample marked as a similar area, and updating, adjusting and obtaining a new set which contains k neighbors and is dissimilar to the area of each sample marked as dissimilar to the area;
an update region weights and feature weights submodule for performing the following operations,
let S be the similarity between the query image p and the image of the ith feedback sample in the set of feedback samplesa(p,gi) The following calculation formula is adopted,
wherein,representing the characteristic weight of the jth region part under the m-dimensional characteristic vector;representing the feature vector of the jth region position in the mth dimension; wj(p,gi) Representing a jth region-based region weight between the query image p and the feedback sample; therein is provided withRepresenting the similarity between the query image p and the feedback sample based on the jth region part under the mth dimension characteristic value;
by using the distance measurement method in machine learning, the following calculation formula is adopted to update the region weight,
Wj(p,gi)＝Wj(p,gi)×β1β1>1 formula II
Wj(p,gi)＝Wj(p,gi)×β20<β2<1 formula III
Wherein, Wj(p,gi) Area weight representing jth area position, β1、β2Is a preset coefficient;
by using the distance measurement method in machine learning, the following calculation formula is adopted for updating the characteristic weight,
wherein, mum、σmRespectively representing the mean value and the variance of the characteristic value of the mth dimension, wherein α is a preset parameter;
the query matching module is used for carrying out feature expression and distance measurement according to the adjusted region weight and the formula I obtained by the weight module to obtain a query matching result;
the result display module is used for displaying the query matching result obtained by the query matching module, and outputting the result if the query matching result meets the requirement; if not, the feedback sample collection submodule is notified to update the feedback sample set until the requirements are met.
The invention can improve the effectiveness and the accuracy of matching identification of the pedestrian re-identification system, and the innovation points mainly comprise the following two points:
1. applying a local area of the image as a feedback unit instead of the whole image to a pedestrian re-identification system, wherein each part corresponds to a weight value;
2. the weight values of each part and the corresponding characteristics are adjusted by using a dynamic K neighbor rule method, and then the sorting result is optimized and improved according to the total similarity value.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention.
Detailed Description
The invention provides a pedestrian re-identification method based on region correlation feedback, which is characterized in that aiming at the remarkable difference between a pedestrian image in a video monitoring scene and various images under other conditions, the pedestrian re-identification method is divided into a plurality of fixed composition regions according to the composition structure characteristics of each part of a body of a person in a walking state, the regions are used as basic processing units, the adjusted dynamic weight value is timely updated and fed back according to the difference of characteristic information contained in each region and the influence of different regions on a final matching result, and finally reordering is carried out according to the optimized similarity, so that the whole process of the region-based correlation feedback method is realized. The specific implementation can adopt a computer software technology to support the process operation. Referring to fig. 1, the flow of the embodiment includes the steps of:
step S1, primary query matching and feedback sample collection: according to the division of the regions, samples with similar regions and dissimilar regions are marked, and the specific process of feedback information collection is realized.
The following substeps may be employed:
step S1.1, performing initial query matching, wherein the initial query matching comprises inputting a target person image as a query image, such as a suspected target person image related to criminal investigation and solution, performing initial query and outputting an initial query sorting result, and when the initial query matching is specifically implemented, the initial query matching of the query image can be performed by adopting a traditional method, such as a Euclidean distance or an L1 distance method, and the query matching results are arranged from high to low according to similarity to obtain an initial sorting result;
and S1.2, feedback information collection is carried out, and when the step S1.2 is executed for the first time, irrelevant images are selected from a certain preset number of images with the top ranking from the primary ranking result to serve as feedback samples and are marked with types to form a feedback sample set. Here, the irrelevant image means that the pedestrian image is not the image of the inquired target person even if the appearances are very similar. Let us divide U regions 1, 2 … U, each feedback sample being respectively labeled as one of 2U types, i.e., similar to the query image based on region 1, dissimilar based on region 1, similar based on region 2, dissimilar based on region 2 …, similar based on region U, dissimilar based on region U. In subsequent iterations, if the current query ranking result of a certain round of execution to the step S4 does not meet the requirement, when returning to the step S1.2, selecting an irrelevant image from a certain preset number of images ranked at the top as a feedback sample and marking the type, adding the irrelevant image into the feedback sample set, and then repeating the operations of the steps S2-S4 based on the expanded feedback sample.
The embodiment is divided into two regions of a trunk part and a leg part. The type of each feedback sample is labeled as a torso-part-based similar sample (PSN)t) Based on torso-part dissimilar samples (PDN)t) Based on similar samples (PSN) at leg partsl) Or based on leg-part dissimilar samples (PDN)l)。
In particular implementation, for the purpose of improving efficiency, the user can select the image of the non-query target as an irrelevant image and mark the irrelevant image according to visual and prior knowledge. For the convenience of user selection reference, similarity measurement can be automatically performed based on the regions and the result is output, and the measurement implementation mode can be self-designated by the person skilled in the art. For example, the pedestrian body structure is divided into an upper torso part (torso part) and a lower leg part (legpart) according to the body structure composition characteristics, and the appearance shape and color features included in the vision are extracted from each region according to the division of the two regions, and an M-dimensional feature vector is obtained, wherein any dimension is defined as the M-th dimension.
Setting a query image p and an unmarked sample library as G, wherein: g ═ GiI ═ 1, …, n }, where n represents the number of samples in the sample library. The marked feedback sample set has n feedback samples. When the feedback sample is divided according to the regions and the similarity comparison is respectively carried out on the feedback sample and the corresponding regions in the query image, the similarity measurement adopted by each region is calculated as the following formula,
wherein,representing the characteristic weight of the jth region part, adopting a preset initial value when the step S1.2 is executed for the first time, and executing the weight value updated in the step S2 when the previous iteration is adopted when the step S1.2 is executed subsequently;an m-dimensional feature vector representing a jth region location;representing a query image p and feedback samples giBased on the similarity of the jth region position under the characteristic value of the mth dimension.
In the embodiment, in an unrelated image class, according to the division of a trunk and legs, similarity comparison is respectively performed on areas corresponding to suspected targets in an inquiry image, and similarity measurement calculation methods respectively adopted by the trunk and the legs are as follows:
andthe characteristic weights respectively represent the leg parts and the trunk parts, the initial value adopted in the step S1.2 is automatically preset by a user according to the similarity of specific different characteristics between the pedestrian image and the feedback sample, and the weight value updated in the step S2 in the previous iteration can be adopted in the subsequent step S1.2;andrespectively representing the m-dimensional characteristic vectors of the leg part and the trunk part;andrespectively representing the similarity between the suspected target and the sample based on the leg region and the trunk part under the mth dimension characteristic value; st(p,gi) The trunk parts are similarValue of the sexual metric, Sl(p,gi) Is the value of the leg portion similarity measure.
After a primary sorting result is obtained according to the similarity metric value, through man-machine interaction, a user can mark a sample (PSN) based on torso part similarity according to visual and prior judgmentt) And dissimilar samples (PDN)t) And based on similar samples (PSN) of leg partsl) And dissimilar samples (PDN)l)。
Step S2, including determining neighbor set, region weight adjustment, feature weight adjustment:
in the step, a similar sample set based on similar regions and dissimilar regions is found out by a region K neighbor set method, and for each region, a pair of sets can be found out; and for each pair of found sets, respectively re-determining new K neighbors by using a dynamic K neighbor method, and adjusting weights corresponding to different areas. For the same region, because the same region contains many pieces of feature information, some features can be obviously distinguished from other samples, the final matching result is greatly influenced and has a larger weight value, and other features are not obviously distinguished from other samples, and the final matching result is slightly influenced and has a smaller weight. To this end, the embodiment of the present invention first finds the neighbor set, and then adjusts the weight. Aiming at a set of dissimilar neighbors in a region, finding out all feature information in the region, expressing all features to form a multi-dimensional feature vector, determining the influence of the feature value on a matching result according to the variation of the feature value corresponding to different dimensions for updated K neighbors, and giving greater weight to features which can well represent the common property of a query target; on the contrary, the characteristics which can not well represent the query target property are endowed with smaller weight values, so that the characteristic weight of different characteristics is adjusted.
The following substeps may be employed:
s2.1, for the query image, firstly, a sample set based on region similarity and region dissimilarity of region similarity is found out by a region K neighbor set method; then, a dynamic k neighbor rule is applied, each sample marked as a similar area is updated, adjusted and obtained to a new similar area set containing k neighbors, and each sample marked as a dissimilar area is updated, adjusted and obtained to a new dissimilar area set containing k neighbors. The specific implementation method comprises the following steps:
for each region, according to the formula (1), a sequence based on the similarity of the region can be obtained, and after the query image is searched by a region K neighbor set method, the query image is respectively searched by usingAndthe representatives form a pair of neighbor sets based on leg similarity and dissimilarity, using SetPSNtAnd SetPDNtRepresenting a pair of neighbor sets composed based on torso similarity and dissimilarity; the K samples included in the neighbor sets are obtained from the feedback samples by a K-neighbor rule, and the value of K may be preset, and in the embodiment, the value of K is taken as K-5.
For each sample marked as similar to a certain region, taking the neighboring set with dissimilar region as a boundary basis, and sequentially selecting samples from the front to the back of the sequencing result until the selected sample belongs to the neighboring set with dissimilar region, so that a new set with similar regions containing k neighbors can be obtained by the method; similarly, for each sample marked as a region dissimilar, a neighboring set similar to the region is used as a boundary basis, and a new set containing k neighbors dissimilar to the region can be obtained by applying the same method. And the k value is dynamically adjusted according to the selected condition of the specific boundary.
Step S2.2, updating the region weight value and the characteristic weight:
let S be the similarity between the query image p and the image of the ith feedback sample in the set of feedback samplesa(p,gi) The following calculation formula is adopted,
wherein,representing the characteristic weight of the jth region part under the m-dimensional characteristic vector;representing the feature vector of the jth region position in the mth dimension; wj(p,gi) Representing a jth region-based region weight between the query image p and the feedback sample; therein is provided withRepresenting the similarity between the query image p and the feedback sample based on the jth region part under the mth dimension characteristic value; j takes the values of 1 and 2 … U,
at step S2.2.1, using the distance measurement method in machine learning, the region weight is updated by the following calculation formula,
Wj(p,gi)＝Wj(p,gi)×β1β1>1 (5)
Wj(p,gi)＝Wj(p,gi)×β20<β2<1 (6)
wherein, Wj(p,gi) Area weight representing jth area position, β1、β2Is a preset coefficient;
at step S2.2.2, using the distance measurement method in machine learning, the updated feature weight uses the following calculation formula,
wherein, mum、σmRespectively representing the mean and variance of the characteristic value of the mth dimension, α is a preset parameter.
In an embodiment of the present invention,
the similarity between the suspected target (image p) and the image of the ith sample in the set of feedback samples can be calculated as follows:
wherein:
andfeature weights under the m-dimensional feature vectors representing the leg and torso parts, respectively;andrespectively representing the m-dimensional characteristic vectors of the leg part and the trunk part; wl(p,gi) And Wt(p,gi) Respectively representing weight values between the suspected target and the sample based on leg regions and trunk part regions;andit represents the similarity between the suspected target and the sample based on the leg region and the torso part under the m-dimensional feature values, respectively.
At step S2.2.1, the region weight values are updated by using the distance measurement method in machine learning,
the region weight values are updated by equations (4) and (5):
Wt,l(p,gi)＝Wt,l(p,gi)×β1β1>1 (9)
Wt,l(p,gi)＝Wt,l(p,gi)×β20<β2<1 (10)
wherein, Wt,l(p,gi) Represents Wt(p,gi) Or Wl(p,gi) When this step is executed for the first time, the previous W is calculatedt(p,gi)、Wl(p,gi) The initial value of (a) can be preset by self according to the difference between the pedestrian query image (image p of the suspected target) selected by the user and the feedback sample in different areas, for example, in some environments, the color of clothes is more vivid, and the appearance difference is not changed much, so that the weight of the upper part of the body is larger. When the step is executed subsequently, the former W is calculatedt(p,gi)、Wl(p,gi) Can be updated according to the calculated weight value when the step is executed in the previous iteration β in the above equation (9)1>1, aiming at samples marked as similar areas, forming new k neighbors closer to the selected query image, and increasing the area weight value of the area; in the above formula (10), 0<β2<1, for the samples with dissimilar regions, the new k neighbors formed by the samples with dissimilar regions are further away from the selected query image, so as to reduce the region weight value of the region, wherein the preset value in the embodiment is a coefficient β1100, factor β20.01. Because new K are obtainedThe neighbor samples are re-selected by the computer according to the dynamic K neighbor rule, so that compared with the original samples, some samples with more obvious characteristics (such as larger body type) in some areas are included in the newly formed set, and accordingly correspond to the query image, and when the samples are matched, the results are greatly influenced by the areas, so that the area weight is changed (β)1Increase)
For neighboring sets with dissimilar regions, the same is true.
At step S2.2.2, the feature weight values are updated by using the distance measurement method in machine learning,
the updated feature weights are adjusted by the following formula:
in equations (11) and (12), the first time this step is performed, it is calculatedAndthe initial value of (a) can be self-preset by a user according to the similarity of the pedestrian image and the feedback sample under different specific characteristics, mum、σmRespectively, the mean and variance of the characteristic value of the mth dimension, and the preset parameter α is generally in the range of 0.5, 1]Internal values, for example 0.5. When the step is executed subsequently, before calculationAndthe value of (c) may be an updated weight value calculated from the previous iteration of performing this step. Obtaining a new region dissimilarity set containing k neighbors after updating, and obtaining a similarity valueAndif the size is larger, the characteristic under the dimension vector cannot truly reflect the actual intention of the user, and the associated matching weight of the characteristic should be reduced; and the values of all similarity sequences are comparedAndstacked to form an Lpn× (2M) matrix, where 2M represents the number of features and the columns of the matrix are the similarity sequencesAndlength L at a certain characteristic valuepnWhen all the area dissimilar samples of the mark have similar similarity values under the characteristic value, the characteristic value represents that the characteristic value well represents the shared property of the query target, so that the reciprocal of the standard deviation of the sequence becomes an unbiased estimation quantity of the characteristic weight, and when the parameter mu ismWhen the value is in the first M/4, the updated feature weight is adjusted by applying the formulas (11) and (12). When the m-dimension characteristic value is reached, a certain image in the samples has a larger similarity with the query image, which does not reflect the real intention of the user (the user aims to find the samples with the characteristic values with larger differences as far as possible, so that the samples are far away from the clustering center as far as possible, the similar samples are ranked forward and close to the clustering center), and therefore the parameter mu is agreed according to experiencemValue is carried under the condition of the front M/4 rangeThe formulas (11) and (12) are used. Mu.sm、σmThe mean and variance of the feature vector composed of features extracted from the newly obtained neighboring sample set (they vary from sample to sample) are used, and thus the feature weights are automatically updated.
Step S3, including feature expression and distance measurement, to obtain the query matching result:
the feature expression and distance metric may be developed using methods well known in the art of pedestrian re-identification.
The feature expression is based on the premise that the features of the image are extracted, and features which can well represent the difference between the image and other samples are selected through training and learning for multiple times on different features through the adjusted weight values.
In the distance measurement, the similarity measurement may be performed again by using the formula (4) through the adjusted weight value obtained in step S2, so as to obtain a new query matching result. Therefore, the area with the increased weight value after adjustment and the characteristic value contained in the area can be fully utilized to find out a sample with higher similarity, and the accuracy of pedestrian re-identification is improved;
step S4, displaying the query matching result obtained in the step S3, and outputting the result if the query matching result meets the requirement; if not, the step S1.2 is returned to iterate until the requirement is met. When the method is specifically implemented, the sequencing result after optimized matching can be output and displayed on an interactive interface, and a user can comprehensively judge and determine whether the query matching result meets the requirements or not by combining the matching target and information of other aspects.
A pedestrian re-recognition system based on region correlation feedback utilizes a correlation feedback technology to optimize a sequencing result, can provide a human-computer interaction interface to facilitate a user to select irrelevant images during specific implementation, marks regions similar or dissimilar to an inquiry image on the images, and utilizes the information to perform correlation feedback. The system can be realized in a software modularization mode and comprises the following modules:
the feedback module is used for carrying out primary query matching and feedback sample collection and comprises the following sub-modules,
the primary query matching sub-module is used for taking the input target person image as a query image, performing initial query and outputting an initial query sequencing result;
the feedback sample collection submodule is used for selecting irrelevant images from a certain preset number of images with the top ranking from the primary ranking result as feedback samples and marking the types to form a feedback sample set when the feedback sample collection is executed for the first time; when the feedback sample collection is subsequently executed, irrelevant images are selected from the query sequencing results obtained by the result display module in the previous iteration as feedback samples, the types of the irrelevant images are marked, and the irrelevant images are added into a feedback sample set;
the marking type mode is that the method is divided into U areas 1 and 2 … U, each feedback sample is marked as one of 2U types, namely similar to the query image based on the area 1, dissimilar based on the area 1, similar based on the area 2, … dissimilar based on the area U, and dissimilar based on the area U; when a feedback sample is marked, extracting visual features in each region according to region division, setting to obtain an M-dimensional feature vector, wherein any dimension is marked as an M-th dimension, and respectively performing similarity comparison with a corresponding region in a query image according to the feature vector;
the weight module is used for determining a neighbor set, adjusting the regional weight and adjusting the characteristic weight, and comprises a sub-module for determining the neighbor set, a sub-module for searching a sample set based on regional similarity and regional dissimilarity by a regional K neighbor set method for a query image; then, applying a dynamic k neighbor rule, updating, adjusting and obtaining a new set which contains k neighbors and is similar to the area of each sample marked as a similar area, and updating, adjusting and obtaining a new set which contains k neighbors and is dissimilar to the area of each sample marked as dissimilar to the area;
an update region weights and feature weights submodule for performing the following operations,
let S be the similarity between the query image p and the image of the ith feedback sample in the set of feedback samplesa(p,gi) The following calculation formula is adopted,
wherein,representing the characteristic weight of the jth region part under the m-dimensional characteristic vector;representing the feature vector of the jth region position in the mth dimension; wj(p,gi) Representing a jth region-based region weight between the query image p and the feedback sample; therein is provided withRepresenting the similarity between the query image p and the feedback sample based on the jth region part under the mth dimension characteristic value;
by using the distance measurement method in machine learning, the following calculation formula is adopted to update the region weight,
Wj(p,gi)＝Wj(p,gi)×β1β1>1 (5)
Wj(p,gi)＝Wj(p,gi)×β20<β2<1 (6)
wherein, Wj(p,gi) Area weight representing jth area position, β1、β2Is a preset coefficient;
by using the distance measurement method in machine learning, the following calculation formula is adopted for updating the characteristic weight,
wherein, mum、σmRespectively representing the mean value and the variance of the characteristic value of the mth dimension, wherein α is a preset parameter;
the query matching module is used for carrying out feature expression and distance measurement according to the adjusted region weight obtained by the weight module and the formula (4) to obtain a query matching result;
the result display module is used for displaying the query matching result obtained by the query matching module, and outputting the result if the query matching result meets the requirement; if not, the feedback sample collection submodule is notified to update the feedback sample set until the requirements are met.
The specific implementation of each module refers to each step of the method flow, and the present invention is not repeated.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.
Claims (7)
1. A pedestrian re-identification method based on region-related feedback is characterized by comprising the following steps: comprises the following steps of (a) carrying out,
step S1, performing initial query matching and feedback sample collection, including the following sub-steps,
step S1.1, performing primary query matching, including taking an input target person image as a query image, performing initial query and outputting an initial query sorting result;
step S1.2, feedback sample collection is carried out, and when the step S1.2 is executed for the first time, irrelevant images are selected from a certain preset number of images which are ranked the most front from the primary sequencing result to serve as feedback samples and mark types to form a feedback sample set; when step S1.2 is executed subsequently, an irrelevant image is selected as a feedback sample from the query ranking result obtained by the previous iteration of step S4, and the type is marked, and a feedback sample set is added;
the marking type mode is that the method is divided into U areas 1 and 2 … U, each feedback sample is marked as one of 2U types, namely similar to the query image based on the area 1, dissimilar based on the area 1, similar based on the area 2, … dissimilar based on the area U, and dissimilar based on the area U; when a feedback sample is marked, extracting visual features in each region according to region division, setting to obtain an M-dimensional feature vector, wherein any dimension is marked as an M-th dimension, and respectively performing similarity comparison with a corresponding region in a query image according to the feature vector;
step S2, determining neighbor set, adjusting region weight and adjusting feature weight, includes the following sub-steps,
s2.1, for the query image, firstly, a sample set based on region similarity and region dissimilarity of region similarity is found out by a region K neighbor set method; then, applying a dynamic k neighbor rule, updating, adjusting and obtaining a new set which contains k neighbors and is similar to the area of each sample marked as a similar area, and updating, adjusting and obtaining a new set which contains k neighbors and is dissimilar to the area of each sample marked as dissimilar to the area;
step S2.2, updating the region weight and the feature weight,
let S be the similarity between the query image p and the image of the ith feedback sample in the set of feedback samplesa(p,gi) The following calculation formula is adopted,
wherein,represents the jth zoneFeature vector of domain position in m dimensionA characteristic weight of; wj(p,gi) Representing a jth region-based region weight between the query image p and the feedback sample; therein is provided withRepresenting characteristic values between the query image p and the feedback sample in the m-dimension based on the j-th region position(ii) similarity of; j takes the values of 1 and 2 … U,
at step S2.2.1, using the distance measurement method in machine learning, the region weight is updated by the following calculation formula,
Wj(p,gi)＝Wj(p,gi)×β1β1>1 formula II
Wj(p,gi)＝Wj(p,gi)×β20<β2<1 formula III
Wherein, Wj(p,gi) Area weight representing jth area position, β1、β2Is a preset coefficient;
at step S2.2.2, using the distance measurement method in machine learning, the updated feature weight uses the following calculation formula,
wherein, mum、σmRespectively representing the mean value and the variance of the characteristic value of the mth dimension, wherein α is a preset parameter;
step S3, according to the adjusted region weight and the formula I obtained in the step S2, feature expression and distance measurement are carried out, and a query matching result is obtained;
step S4, displaying the query matching result obtained in the step S3, and outputting the result if the query matching result meets the requirement; if not, the step S1.2 is returned to iterate until the requirement is met.
2. The pedestrian re-identification method based on the region-dependent feedback according to claim 1, wherein: when the feedback sample is divided according to the regions and the similarity comparison is respectively carried out on the feedback sample and the corresponding regions in the query image, the similarity measurement adopted by each region is calculated as the following formula,
formula five
Wherein,the characteristic weight of the jth region part is represented, the step S1.2 is executed for the first time by adopting a preset initial value, and the step S1.2 is executed when the previous iteration is adopted in the subsequent stepThe updated weight value of S2;an m-dimension feature vector representing a j-th region position;representing a query image p and feedback samples giBased on the jth region position in the mth dimension characteristic valueThe following similarity.
3. The pedestrian re-identification method based on the region-dependent feedback according to claim 2, wherein: when 2 regions are divided, including the trunk and the legs, and similarity comparison is respectively carried out on the regions corresponding to the trunk and the legs in the query image, similarity measurement adopted by the trunk and the legs respectively is calculated as the following formula,
wherein,andrepresenting the characteristic weights of the leg and torso parts respectively,andrespectively represent the legAnd a trunk part m-dimensional feature vector;andrespectively representing the characteristic value of the m-dimension between the query image and the feedback sample based on the legDegree of similarity of lower body and m-dimension characteristic value based on body part(ii) similarity of; st(p,gi) Is the value of the torso-part similarity measure, Sl(p,gi) Is the value of the leg portion similarity measure.
4. The pedestrian re-identification method based on the region-dependent feedback according to claim 3, wherein: the similarity between the query image p and the image of the ith sample in the feedback sample set is calculated as follows,
wherein,andfeature weights under the m-dimensional feature vectors representing the leg and torso parts, respectively;andrespectively representing the m-dimensional characteristic vectors of the leg part and the trunk part; wl(p,gi) And Wt(p,gi) Respectively representing leg region and trunk region based weights between the suspect target and the sample;andrespectively representing the characteristic values of the m-dimension between the suspected target and the sample based on the leg regionsDegree of similarity of lower body and m-dimension characteristic value based on body partThe following similarity.
5. The pedestrian re-identification method based on the region-dependent feedback according to claim 4, wherein: at step S2.2.1, using the distance measurement method in machine learning, the region weight is updated by the following calculation formula,
Wt,l(p,gi)＝Wt,l(p,gi)×β1β1>1 formula nine
Wt,l(p,gi)＝Wt,l(p,gi)×β20<β2<Formula 1 ten
Wherein, Wt,l(p,gi) Represents Wt(p,gi) Or Wl(p,gi)，β1、β2Is a preset coefficient.
6. The pedestrian re-identification method based on the region-dependent feedback according to claim 4, wherein: at step S2.2.2, using the distance measurement method in machine learning, the updated feature weight uses the following calculation formula,
wherein, mum、σmRespectively representing the mean and variance of the characteristic value of the mth dimension, α is a preset parameter.
7. A pedestrian re-identification system based on region-dependent feedback is characterized in that: comprises the following modules which are used for realizing the functions of the system,
the feedback module is used for carrying out primary query matching and feedback sample collection and comprises the following sub-modules,
the primary query matching sub-module is used for taking the input target person image as a query image, performing initial query and outputting an initial query sequencing result;
the feedback sample collection submodule is used for selecting irrelevant images from a certain preset number of images with the top ranking from the primary ranking result as feedback samples and marking the types to form a feedback sample set when the feedback sample collection is executed for the first time; when the feedback sample collection is subsequently executed, irrelevant images are selected from the query sequencing results obtained by the result display module in the previous iteration as feedback samples, the types of the irrelevant images are marked, and the irrelevant images are added into a feedback sample set;
the marking type mode is that the method is divided into U areas 1 and 2 … U, each feedback sample is marked as one of 2U types, namely similar to the query image based on the area 1, dissimilar based on the area 1, similar based on the area 2, … dissimilar based on the area U, and dissimilar based on the area U; when a feedback sample is marked, extracting visual features in each region according to region division, setting to obtain an M-dimensional feature vector, wherein any dimension is marked as an M-th dimension, and respectively performing similarity comparison with a corresponding region in a query image according to the feature vector;
the weight module is used for determining a neighbor set, adjusting the regional weight and adjusting the characteristic weight, and comprises a sub-module for determining the neighbor set, a sub-module for searching a sample set based on regional similarity and regional dissimilarity by a regional K neighbor set method for a query image; then, applying a dynamic k neighbor rule, updating, adjusting and obtaining a new set which contains k neighbors and is similar to the area of each sample marked as a similar area, and updating, adjusting and obtaining a new set which contains k neighbors and is dissimilar to the area of each sample marked as dissimilar to the area;
an update region weights and feature weights submodule for performing the following operations,
let S be the similarity between the query image p and the image of the ith feedback sample in the set of feedback samplesa(p,gi) The following calculation formula is adopted,
wherein,representing the feature vector of the j-th region in m dimensionA characteristic weight of; wj(p,gi) Representing a jth region-based region weight between the query image p and the feedback sample; therein is provided withRepresenting characteristic values between the query image p and the feedback sample in the m-dimension based on the j-th region position(ii) similarity of;
by using the distance measurement method in machine learning, the following calculation formula is adopted to update the region weight,
Wj(p,gi)＝Wj(p,gi)×β1β1>1 formula II
Wj(p,gi)＝Wj(p,gi)×β20<β2<1 formula III
Wherein, Wj(p,gi) Area weight representing jth area position, β1、β2Is a preset coefficient;
by using the distance measurement method in machine learning, the following calculation formula is adopted for updating the characteristic weight,
wherein, mum、σmRespectively representing the mean value and the variance of the characteristic value of the mth dimension, wherein α is a preset parameter;
the query matching module is used for carrying out feature expression and distance measurement according to the adjusted region weight and the formula I obtained by the weight module to obtain a query matching result;
the result display module is used for displaying the query matching result obtained by the query matching module, and outputting the result if the query matching result meets the requirement; if not, the feedback sample collection submodule is notified to update the feedback sample set until the requirements are met.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410076028.3A CN103793721B (en) | 2014-03-04 | 2014-03-04 | Pedestrian repeat recognition method and system based on area related feedback |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410076028.3A CN103793721B (en) | 2014-03-04 | 2014-03-04 | Pedestrian repeat recognition method and system based on area related feedback |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103793721A CN103793721A (en) | 2014-05-14 |
| CN103793721B true CN103793721B (en) | 2017-05-10 |
Family
ID=50669363
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410076028.3A Expired - Fee Related CN103793721B (en) | 2014-03-04 | 2014-03-04 | Pedestrian repeat recognition method and system based on area related feedback |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103793721B (en) |
Families Citing this family (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104200206B (en) * | 2014-09-09 | 2017-04-26 | 武汉大学 | Double-angle sequencing optimization based pedestrian re-identification method |
| CN104376334B (en) * | 2014-11-12 | 2018-05-29 | 上海交通大学 | A kind of pedestrian comparison method of multi-scale feature fusion |
| CN104376212B (en) * | 2014-11-17 | 2016-12-21 | 深圳市银雁金融服务有限公司 | The method and device of assessment operation accuracy |
| CN104462550B (en) * | 2014-12-25 | 2017-07-11 | 武汉大学 | Pedestrian's recognition methods again of sorting consistence is merged based on similitude and dissimilarity |
| CN106557533B (en) * | 2015-09-24 | 2020-03-06 | 杭州海康威视数字技术股份有限公司 | Single-target multi-image joint retrieval method and device |
| CN105488502B (en) * | 2015-11-27 | 2018-12-21 | 北京航空航天大学 | Object detection method and device |
| CN108229521B (en) * | 2017-02-23 | 2020-09-15 | 北京市商汤科技开发有限公司 | Object recognition network training method, device and system and application thereof |
| CN108960013B (en) * | 2017-05-23 | 2020-09-15 | 深圳荆虹科技有限公司 | Pedestrian re-identification method and device |
| CN107563327B (en) * | 2017-08-31 | 2021-07-20 | 武汉大学 | Pedestrian re-identification method and system based on self-walking feedback |
| CN109697391A (en) * | 2017-10-23 | 2019-04-30 | 北京京东尚科信息技术有限公司 | Personage knows method for distinguishing, system and terminal device again in closing place |
| CN108052665B (en) * | 2017-12-29 | 2020-05-05 | 深圳市中易科技有限责任公司 | Data cleaning method and device based on distributed platform |
| CN108710824A (en) * | 2018-04-10 | 2018-10-26 | 国网浙江省电力有限公司信息通信分公司 | A kind of pedestrian recognition method divided based on regional area |
| CN110909565B (en) * | 2018-09-14 | 2023-06-16 | 阿里巴巴集团控股有限公司 | Image recognition and pedestrian re-recognition method and device, electronic and storage equipment |
| CN109740541B (en) * | 2019-01-04 | 2020-08-04 | 重庆大学 | Pedestrian re-identification system and method |
| CN110377774B (en) * | 2019-07-15 | 2023-08-01 | 腾讯科技（深圳）有限公司 | Method, device, server and storage medium for person clustering |
| CN110536305B (en) * | 2019-08-29 | 2023-09-12 | 武汉赛可锐信息技术有限公司 | WiFi hot spot detection method and device, terminal equipment and storage medium |
| CN111914241B (en) * | 2020-08-06 | 2024-01-30 | 上海熙菱信息技术有限公司 | Method for dynamically identifying unstructured object identity information |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101539930A (en) * | 2009-04-21 | 2009-09-23 | 武汉大学 | Search method of related feedback images |
| CN102663366A (en) * | 2012-04-13 | 2012-09-12 | 中国科学院深圳先进技术研究院 | Method and system for identifying pedestrian target |
| CN103325122A (en) * | 2013-07-03 | 2013-09-25 | 武汉大学 | Pedestrian retrieval method based on bidirectional sequencing |
-
2014
- 2014-03-04 CN CN201410076028.3A patent/CN103793721B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101539930A (en) * | 2009-04-21 | 2009-09-23 | 武汉大学 | Search method of related feedback images |
| CN102663366A (en) * | 2012-04-13 | 2012-09-12 | 中国科学院深圳先进技术研究院 | Method and system for identifying pedestrian target |
| CN103325122A (en) * | 2013-07-03 | 2013-09-25 | 武汉大学 | Pedestrian retrieval method based on bidirectional sequencing |
Non-Patent Citations (3)
| Title |
|---|
| "Interactive person re-identification in TV series";Fischer M. et al.;《2010 International Workshop on Context-based Multimedia Indexing》;20101231;第1-6页 * |
| "POP: Person Re-Identification Post-Rank Optimisation";Chunxiao Liu et al.;《Proceeding of the IEEE International Conference on Computer Vision》;20131231;第441-448页 * |
| "Relevance Feedback for CBIR: A New Approach Based on Probabilistic Feature Weighting With Positive and Negative Examples";Kherfi M.L. et al.;《IEEE TRANSACTIONS ON IMAGE PROCESSING》;20060430;第15卷(第4期);第1017-1030页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103793721A (en) | 2014-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103793721B (en) | Pedestrian repeat recognition method and system based on area related feedback | |
| CN111709311B (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| CN106033435B (en) | Item identification method and device, indoor map generation method and device | |
| CN103778227B (en) | The method screening useful image from retrieval image | |
| Cui et al. | Target transformed regression for accurate tracking | |
| Jiang et al. | Online robust action recognition based on a hierarchical model | |
| CN103336835B (en) | Image retrieval method based on weight color-sift characteristic dictionary | |
| Zhang et al. | A survey on human pose estimation | |
| Hu et al. | Exemplar-based recognition of human–object interactions | |
| CN111046715B (en) | Human body action contrast analysis method based on image retrieval | |
| CN107330363B (en) | Rapid internet billboard detection method | |
| Liao et al. | Ai golf: Golf swing analysis tool for self-training | |
| CN110532409B (en) | Image retrieval method based on heterogeneous bilinear attention network | |
| CN109325407A (en) | Optical remote sensing video object detection method based on F-SSD network filtering | |
| CN113792686B (en) | Vehicle re-identification method based on visual representation of invariance across sensors | |
| CN109784295B (en) | Video stream feature identification method, device, equipment and storage medium | |
| Pang et al. | Analysis of computer vision applied in martial arts | |
| US8724890B2 (en) | Vision-based object detection by part-based feature synthesis | |
| CN107563327B (en) | Pedestrian re-identification method and system based on self-walking feedback | |
| CN108845999A (en) | A kind of trademark image retrieval method compared based on multiple dimensioned provincial characteristics | |
| CN117315556A (en) | Improved Vision Transformer insect fine grain identification method | |
| Ran et al. | Improving visible-thermal ReID with structural common space embedding and part models | |
| CN111797704A (en) | Action recognition method based on related object perception | |
| Wang et al. | Occluded person re-identification based on differential attention siamese network | |
| CN111144469A (en) | End-to-end multi-sequence text recognition method based on multi-dimensional correlation time sequence classification neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170510 |