CN111967408B - Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification - Google Patents

Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification Download PDF

Info

Publication number
CN111967408B
CN111967408B CN202010843411.2A CN202010843411A CN111967408B CN 111967408 B CN111967408 B CN 111967408B CN 202010843411 A CN202010843411 A CN 202010843411A CN 111967408 B CN111967408 B CN 111967408B
Authority
CN
China
Prior art keywords
resolution
picture
low
super
scale
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
Application number
CN202010843411.2A
Other languages
Chinese (zh)
Other versions
CN111967408A (en
Inventor
王亮
黄岩
韩苛
单彩峰
纪文峰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Cas Artificial Intelligence Research Qingdao Co ltd
Original Assignee
Cas Artificial Intelligence Research Qingdao Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Cas Artificial Intelligence Research Qingdao Co ltd filed Critical Cas Artificial Intelligence Research Qingdao Co ltd
Priority to CN202010843411.2A priority Critical patent/CN111967408B/en
Publication of CN111967408A publication Critical patent/CN111967408A/en
Application granted granted Critical
Publication of CN111967408B publication Critical patent/CN111967408B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformation in the plane of the image
    • G06T3/40—Scaling the whole image or part thereof
    • G06T3/4053—Super resolution, i.e. output image resolution higher than sensor resolution
    • G06T3/4076—Super resolution, i.e. output image resolution higher than sensor resolution by iteratively correcting the provisional high resolution image using the original low-resolution image

Abstract

The invention discloses a low-resolution pedestrian re-identification method and a system based on 'prediction-recovery-identification', comprising the following steps: inputting the acquired low-resolution picture to be identified into a trained deep neural network model, and performing detail recovery under the optimal scale to obtain a super-resolution picture; and calculating Euclidean distance between the super-resolution picture features and the high-resolution search library picture features, and sequencing identity matching according to the Euclidean distance. The predictor of the deep neural network model can adaptively predict a better scale factor according to the content of the low-resolution picture so as to achieve better recovery and recognition effects.

Description

Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification
Technical Field
The invention relates to the technical field of pattern recognition and machine learning, in particular to a low-resolution pedestrian re-recognition method and system based on 'prediction-recovery-recognition'.
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
The goal of pedestrian re-identification is to give a pedestrian query picture (query) under one camera and match the picture of the same pedestrian in the picture library (galery) of another camera. With the wider and wider distribution of the cameras of the street alleys, the pedestrian re-identification shows wide application prospect in the aspects of security protection and the like. For example, scenes such as finding lost children and tracking suspects can be assisted by a pedestrian re-identification technology.
Low resolution pedestrian re-identification refers to a case where the resolution of a pedestrian query picture (query) is low, and the resolution of pictures in a picture library (galery) is high. Many previous methods use super-resolution modules for picture enlargement and detail restoration of low resolution pictures. However, they often preset a fixed scale factor when the image is over-divided, which easily causes problems of insufficient detail recovery, or excessive noise generation, and the like, and it is difficult to ensure that the recovery effect of the image is most suitable for identifying the identity.
Disclosure of Invention
In order to solve the above problems, the invention provides a low-resolution pedestrian re-identification method and system based on 'prediction-recovery-identification', which can adaptively predict a better scale factor according to the content of a low-resolution picture so as to achieve better recovery and identification effects.
In some embodiments, the following technical scheme is adopted:
the low-resolution pedestrian re-identification method based on prediction-recovery-identification comprises the following steps:
inputting the acquired low-resolution picture to be identified into a trained deep neural network model, and performing detail recovery under the optimal scale to obtain a super-resolution picture;
and calculating Euclidean distance between the super-resolution picture features and the high-resolution search library picture features, and sequencing identity matching according to the Euclidean distance.
In other embodiments, the following technical solutions are adopted:
a "prediction-restoration-recognition" based low resolution pedestrian re-recognition system comprising:
the device is used for inputting the acquired low-resolution picture to be recognized into the trained deep neural network model, and performing detail recovery under the optimal scale to obtain a super-resolution picture;
and the device is used for calculating the Euclidean distance between the super-resolution picture features and the high-resolution search library picture features and sequencing identity matching according to the Euclidean distance.
In other embodiments, the following technical solutions are adopted:
a terminal device comprising a processor and a computer-readable storage medium, the processor being configured to implement instructions; the computer readable storage medium is for storing a plurality of instructions adapted to be loaded by a processor and to perform the above-described "prediction-restoration-recognition" based low resolution pedestrian re-recognition method.
In other embodiments, the following technical solutions are adopted:
a computer readable storage medium having stored therein a plurality of instructions adapted to be loaded by a processor of a terminal device and to execute the above-mentioned "prediction-restoration-recognition" based low resolution pedestrian re-recognition method.
Compared with the prior art, the invention has the beneficial effects that:
(1) the predictor of the deep neural network model can adaptively predict a better scale factor according to the content of the low-resolution picture so as to achieve better recovery and recognition effects.
(2) The scale factor measurement provided by the invention evaluates the quality of each scale factor in a self-supervision mode without depending on manual marking of the optimal scale factor, and the whole network model can be trained end to end, thereby improving the training efficiency.
Drawings
Fig. 1 is a flowchart of a low-resolution pedestrian re-identification method based on "prediction-recovery-identification" in the embodiment of the present invention.
Detailed Description
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
Example one
In one or more embodiments, a low-resolution pedestrian re-identification method based on "prediction-recovery-identification" is disclosed, which can predict a better scale factor for a low-resolution pedestrian picture, perform detail recovery and picture amplification under the scale factor, and then perform identity identification on the recovered picture.
The method specifically comprises the following steps:
(1) inputting the acquired low-resolution picture to be identified into a trained deep neural network model, and performing detail recovery under the optimal scale to obtain a super-resolution picture;
(2) and calculating Euclidean distance between the super-resolution picture features and the high-resolution search library picture features, and sequencing identity matching according to the Euclidean distance.
It should be noted that the resolution in this embodiment refers to the product of the picture width and the picture height; the resolution is not absolute, but rather a relative concept, the resolution of the picture in the high resolution search library (gallery) is typically 2 times to 4 times higher than the query picture (query) (i.e. the low resolution picture to be identified), or even more. When the super-resolution restoration is performed on the picture, it is often necessary to specify a scale factor, for example, the scale factor is 3, which means that the super-resolution picture is obtained after the width and the height of the picture are both amplified by 3 times.
Specifically, in this embodiment, the process of establishing the deep neural network model and training the model specifically includes the following steps:
(1) and establishing a deep neural network model and setting a corresponding module network structure. As shown in FIG. 1, the model comprises three parts, namely an adaptive Scale Factor predictor P (adaptive Scale Factor predictor), a Super-Resolution Module G (Super-Resolution Module) and a Re-identification Module F (Re-id Module). In addition, N selectable scale factors { r } are preset1,r2,…,rN}。
(2) Grouping data in a training data set, one set of training data consisting of three pictures { xl,xsl,xhAnd (9) composition. x is the number oflIs a low resolution picture, xhIs a high resolution picture with the same identity tag but a different camera tag. x is the number ofslIs formed by xhThe resulting composite low resolution picture is directly downsampled.
(3) X is to belInput to the predictor P. P can be viewed as a classifier, outputting xlRespectively belonging to N classes r1,r2,…,rNProbability of
Figure BDA0002642234490000041
Can be expressed as
Figure BDA0002642234490000042
Figure BDA0002642234490000043
Also denotes xlRespectively { r }1,r2,…,rNProbability of the next frame. The scale factor corresponding to the maximum probability is the predicted optimal scale rpI.e. by
Figure BDA0002642234490000044
(4) X is to belAnd xslAnd inputting the image data into a super-resolution module G for detail recovery and image amplification. x is the number oflIn turn with { r1,r2,…,rNMagnifying the scale factors to obtain N corresponding super-resolution pictures
Figure BDA0002642234490000051
xslRandomly selecting a scale factor rsl∈{r1,r2,…,rNAmplifying to obtain corresponding super-resolution picture
Figure BDA0002642234490000052
High resolution picture xhAnd super-resolution pictures
Figure BDA0002642234490000053
Form a super-resolution loss LsrAnd the method is used for training the super-resolution module. L issrCan be expressed as
Figure BDA0002642234490000054
Wherein Xi,jIs the pixel value on the picture X coordinate (i, j), and W and H are the width and height of the picture, respectively.
(5) All super-resolution pictures
Figure BDA0002642234490000055
And high resolution picture xhSending the pictures into a re-identification module F to extract features, and sequentially representing the features of the pictures as
Figure BDA0002642234490000056
To compare xlIn the corresponding pictures with different scales, which recovers more discriminative content, we propose the concept of scale factor measurement. Scale factor measurement in feature space
Figure BDA0002642234490000057
Regarded as "anchors", calculated in turn
Figure BDA0002642234490000058
And
Figure BDA0002642234490000059
the euclidean distance of (c). When evaluating the scale factors, the probability that the scale factors corresponding to the features with closer distances are the optimal factors is higher.
The specific formula is as follows:
Figure BDA00026422344900000510
wherein d isiIs that
Figure BDA00026422344900000511
And
Figure BDA00026422344900000512
the euclidean distance between them, gamma, is a regulatory factor.
Figure BDA00026422344900000513
Is represented by riIs a low resolution picture xlThe probability of the best scale truth value. Gamma is an odd number to ensure a distance x in the feature spacehCloser scale factors are given higher probability.
We use
Figure BDA00026422344900000514
Forming an N-dimensional real valued vector
Figure BDA00026422344900000515
Called dynamic soft label, for supervising the prediction result of predictor P
Figure BDA00026422344900000516
Defining a predicted loss LpThe following:
Figure BDA0002642234490000061
in use error inverseIn the case of the forward propagation algorithm, we cut off the error from LpPropagation to dynamic soft tags. That is, the evaluated dynamic soft label is regarded as a true value and is used only for the supervision predictor P, and the super-resolution module G and the re-recognition module F are not included.
(6) Defining identity cross-entropy loss L using identity tag data in a datasetidAnd identity ternary loss Ltrip. The total loss L when training the model is as follows:
L＝Lsr+αLid+βLtrip+λLp
alpha, beta and lambda are all weight factors. We use an error back-propagation algorithm to minimize the objective function and thus optimize the parameters of the model.
(7) In order to give better initialization weights to the super-resolution module G and the re-identification module F, we perform pre-training before training the entire network. The predictor P and scale factor metrics are removed during pre-training. x is the number oflAnd xslA scale factor is randomly selected, amplified by a super-resolution module G and then input into a re-identification module F, and xhThen directly into F. The predicted loss L is also correspondingly removed from the total loss functionpBy minimizing L onlysr,LidAnd LtripThe optimization of the model parameters is performed.
(8) When testing, inputting low resolution query picture (query) into predictor P, and selecting the scale factor r with highest prediction probabilitypAnd G, as an optimal scale, restoring details according to the scale to obtain a super-resolution picture, and finally inputting the super-resolution picture into a re-identification module F. And the high resolution search library picture (gallery) is directly input to the re-recognition module F. In the feature space, the similarity of the super-resolution picture and the high-resolution picture is measured by calculating the Euclidean distance between the features of the super-resolution picture and the high-resolution picture. The smaller the euclidean distance between two picture features, the higher the similarity. The higher the similarity, the more forward the search pool picture (gallery) is located in the identity matching result of the query picture (query).
For the purpose of illustrating the detailed embodiments of the present invention, the MLR-CUHK03 data set is taken as an example. In this example, predictor P and re-recognition module F are both ResNet50 modules pre-trained on the ImageNet dataset, and a 1 × 1 convolutional layer is added after the global average pooling layer of the ResNet50 module to reduce the dimension of the feature vector from 2014 to 512. Super-resolution module G is a metassr module pre-trained on the DIV2K dataset. And predicting 4 scale factors as {1,2,3 and 4}, wherein the 4 scale factors respectively represent that the picture is amplified by 1 time, 2 times, 3 times and 4 times. The method comprises the following specific steps:
step S1, each low resolution picture x in the training data setlHigh resolution map x randomly identical to an identity tag but different from a camera taghA set of training data is formed. And is xhRandomly selecting a scale factor, and performing down-sampling of bilinear interpolation to obtain a synthesized low-resolution picture xsl。
And step S2, pre-training the super-resolution module G and the re-recognition module F. x is the number oflAnd xslA scale factor is randomly selected, amplified by a super-resolution module G and then input to a re-identification module F, and xhThen directly into F. Loss L at super resolutionsrIdentity cross entropy loss LidAnd identity ternary loss LtripIs the total loss function, and G and P are optimized by minimizing the loss function. The training period for pre-training is set to 20. And the optimized G and P are used as initialization modules when the whole model is trained.
Step S3 to predict loss LpSuper-resolution loss LsrIdentity cross entropy loss LidAnd identity ternary loss LtripThe weighted sum of (a) is the total loss function, and the entire model is trained. The training period is set to 60.
Step S4, during testing, a low-resolution query picture (query) is input into the predictor P, and the probability that each scale factor is the optimal scale of the picture is output. The scale factor r with the highest prediction probabilitypAnd G, as an optimal scale, restoring details according to the scale to obtain a super-resolution picture, and inputting the super-resolution picture into a re-identification module F to extract features. High resolution search library picture (gallery) direct inputTo the re-identification module F. In the feature space, the similarity of the super-resolution picture features and the high-resolution picture features is measured by calculating the Euclidean distance between the super-resolution picture features and the high-resolution picture features. The smaller the euclidean distance between two picture features, the higher the similarity. And sorting the identity matching according to the sequence of the similarity of the characteristic distance between the search library picture (galery) and the query picture (query) from high to low.
Example two
In one or more embodiments, a "prediction-restoration-recognition" based low resolution pedestrian re-recognition system is disclosed, comprising:
the device is used for inputting the acquired low-resolution picture to be recognized into the trained deep neural network model, and performing detail recovery under the optimal scale to obtain a super-resolution picture;
and the device is used for calculating the Euclidean distance between the super-resolution picture features and the high-resolution search library picture features and sequencing identity matching according to the Euclidean distance.
It should be noted that the specific implementation manner of the apparatus is the same as that of the method disclosed in the first embodiment, and is not described again.
EXAMPLE III
In one or more implementations, a terminal device is disclosed, comprising a server including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement a low resolution pedestrian re-identification method based on "prediction-recovery-identification" in the first embodiment. For brevity, no further description is provided herein.
It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general purpose processors, digital signal processors DSP, application specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and so on.
A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The memory may include both read-only memory and random access memory, and may provide instructions and data to the processor, and a portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.
In implementation, the steps of the above method may be performed by integrated logic circuits of hardware in a processor or instructions in the form of software.
The low-resolution pedestrian re-identification method based on the prediction-recovery-identification in the first embodiment can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor. The software modules may be located in ram, flash, rom, prom, or eprom, registers, among other storage media as is well known in the art. The storage medium is located in a memory, and a processor reads information in the memory and completes the steps of the method in combination with hardware of the processor. To avoid repetition, it is not described in detail here.
Although the embodiments of the present invention have been described with reference to the accompanying drawings, it is not intended to limit the scope of the present invention, and it should be understood by those skilled in the art that various modifications and variations can be made without inventive efforts by those skilled in the art based on the technical solution of the present invention.

Claims (9)

1. The low-resolution pedestrian re-identification method based on prediction-recovery-identification is characterized by comprising the following steps of:
inputting the acquired low-resolution picture to be identified into a trained deep neural network model, and performing detail recovery under the optimal scale to obtain a super-resolution picture;
calculating Euclidean distance between the super-resolution picture features and the high-resolution search library picture features, and sequencing identity matching according to the Euclidean distance; the deep neural network model includes:
a predictor for predicting each preset of the low resolution picturesThe probability that the scales are the optimal scales, respectively; specifically, the predictor extracts the characteristics of the low-resolution pictures, and the output low-resolution pictures belong to N categories { r }respectively1，r2，...，rNProbability of (c) } of (c)
Figure FDA0003602567260000011
The described
Figure FDA0003602567260000012
Optimal scale factor r representing low resolution pictures1，r2，...，rNAnd (4) realizing the measurement which is the optimal scale probability for each scale factor through the scale factor measurement, wherein the formula is as follows:
Figure FDA0003602567260000013
wherein N selectable scale factors { r ] are preset1，r2，...，rN}，
Figure FDA0003602567260000014
Represents a scale factor riIs a low resolution picture xlProbability of best scale truth, diRepresenting the Euclidean distance between each super-resolution picture feature corresponding to the low-resolution picture and the picture feature of the high-resolution search library, wherein gamma is a regulation factor;
the super-resolution module is used for traversing the N scale factors of the low-resolution picture and respectively generating N corresponding super-resolution pictures;
and the re-identification module is used for extracting picture features and calculating the Euclidean distance between each super-resolution picture feature corresponding to the low-resolution picture and the picture feature of the high-resolution search library so as to measure the similarity between the two picture features.
2. The "prediction-restoration-recognition-based" low-resolution pedestrian re-recognition method of claim 1, wherein the training process for the deep neural network model comprises:
grouping pictures in the constructed training data set, wherein each group of training data comprises a low-resolution picture, a high-resolution picture with the same identity label but different camera labels, and a synthesized low-resolution picture obtained by down-sampling the high-resolution picture;
for each group of training data sets, predicting the probability that each preset scale of the low-resolution pictures is the optimal scale through a deep neural network model, traversing N scale factors of the low-resolution pictures, and generating N corresponding super-resolution pictures respectively; randomly selecting a scale factor for synthesizing the low-resolution picture to generate 1 super-resolution picture;
calculating the Euclidean distance between each super-resolution picture feature and the high-resolution picture feature, wherein the calculation result is used as a dynamic soft label and represents the evaluation probability that each scale factor is the optimal scale;
different loss functions are constructed, and network parameters are trained and optimized by minimizing a weighted sum of the different loss functions.
3. The "prediction-restoration-recognition-based" low-resolution pedestrian re-recognition method of claim 2, wherein the loss function comprises: a resolution loss function, an identity cross entropy loss function, an identity ternary loss function and a prediction loss function; the dynamic soft label is used as a true value supervision signal of the prediction probability to form a prediction loss function.
4. The method for re-identifying pedestrians in low resolution based on "prediction-recovery-recognition" as claimed in claim 2, wherein, when training with each batch of training data, randomly selecting one from N scale factors for the input low resolution picture and the synthesized low resolution picture, and performing picture recovery in corresponding scale; and removing the predictor and the prediction loss function, taking the weighted sum of the super-resolution loss function, the identity cross entropy loss function and the identity ternary loss function as the total loss, and pre-training the super-resolution module and the re-recognition module by minimizing the total loss.
5. The method for re-identifying pedestrians in low resolution based on "prediction-recovery-recognition" as claimed in claim 2, wherein the tested low resolution picture is inputted into the predictor, the scale factor with the highest expected probability is used as the optimal scale, and then the super resolution picture is recovered according to the scale and sent into the re-identification module; the tested high-resolution picture is directly input to a re-identification module;
and calculating Euclidean distance between the super-resolution picture features and the high-resolution picture features in the feature space to measure the similarity of the super-resolution picture features and the high-resolution picture features.
6. The "prediction-recovery-recognition-based" low-resolution pedestrian re-recognition method according to claim 5, wherein the higher the similarity, the higher the position of the high-resolution search library picture in the identity matching result.
7. A "prediction-restoration-recognition" based low resolution pedestrian re-recognition system, comprising:
the device is used for inputting the acquired low-resolution picture to be identified into the trained deep neural network model, and performing detail recovery under the optimal scale to obtain a super-resolution picture;
the device is used for calculating Euclidean distances between the super-resolution picture features and the high-resolution search library picture features and sequencing identity matching according to the Euclidean distances; the deep neural network model includes:
the predictor is used for predicting the probability that each preset scale of the low-resolution picture is the optimal scale; specifically, the predictor extracts the characteristics of the low-resolution pictures, and the output low-resolution pictures belong to N categories { r }respectively1，r2，...，rNProbability of
Figure FDA0003602567260000031
The above-mentioned
Figure FDA0003602567260000032
Optimal scale factor r representing low resolution pictures1，r2，...，rNAnd (4) realizing the measurement which is the optimal scale probability for each scale factor through the scale factor measurement, wherein the formula is as follows:
Figure FDA0003602567260000033
wherein N selectable scale factors { r ] are preset1，r2，...，rN}，
Figure FDA0003602567260000034
Represents a scale factor riIs a low resolution picture xlProbability of best scale truth, diRepresenting the Euclidean distance between each super-resolution picture feature corresponding to the low-resolution picture and the picture feature of the high-resolution search library, wherein gamma is a regulation factor;
the super-resolution module is used for traversing the N scale factors of the low-resolution picture and respectively generating N corresponding super-resolution pictures;
and the re-identification module is used for extracting picture features and calculating the Euclidean distance between each super-resolution picture feature corresponding to the low-resolution picture and the picture feature of the high-resolution search library so as to measure the similarity between the two picture features.
8. A terminal device comprising a processor and a computer-readable storage medium, the processor being configured to implement instructions; the computer-readable storage medium storing a plurality of instructions adapted to be loaded by a processor and to perform the "prediction-restoration-recognition" based low resolution pedestrian re-recognition method of any one of claims 1 to 6.
9. A computer-readable storage medium having stored therein a plurality of instructions, wherein the instructions are adapted to be loaded by a processor of a terminal device and to perform the "prediction-restoration-recognition" based low resolution pedestrian re-recognition method according to any one of claims 1 to 6.
CN202010843411.2A 2020-08-20 2020-08-20 Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification Active CN111967408B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010843411.2A CN111967408B (en) 2020-08-20 2020-08-20 Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010843411.2A CN111967408B (en) 2020-08-20 2020-08-20 Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification

Publications (2)

Publication Number Publication Date
CN111967408A CN111967408A (en) 2020-11-20
CN111967408B true CN111967408B (en) 2022-06-21

Family

ID=73388631

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010843411.2A Active CN111967408B (en) 2020-08-20 2020-08-20 Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification

Country Status (1)

Country Link
CN (1) CN111967408B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113420697B (en) * 2021-07-01 2022-12-09 中科人工智能创新技术研究院(青岛)有限公司 Reloading video pedestrian re-identification method and system based on appearance and shape characteristics
CN114359113A (en) * 2022-03-15 2022-04-15 天津市电子计算机研究所有限公司 Detection method and application system of face image reconstruction and restoration method

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993072A (en) * 2019-03-14 2019-07-09 中山大学 The low resolution pedestrian weight identifying system and method generated based on super resolution image
CN111275171A (en) * 2020-01-19 2020-06-12 合肥工业大学 Small target detection method based on parameter sharing and multi-scale super-resolution reconstruction
CN111461039A (en) * 2020-04-07 2020-07-28 电子科技大学 Landmark identification method based on multi-scale feature fusion

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11651206B2 (en) * 2018-06-27 2023-05-16 International Business Machines Corporation Multiscale feature representations for object recognition and detection

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109993072A (en) * 2019-03-14 2019-07-09 中山大学 The low resolution pedestrian weight identifying system and method generated based on super resolution image
CN111275171A (en) * 2020-01-19 2020-06-12 合肥工业大学 Small target detection method based on parameter sharing and multi-scale super-resolution reconstruction
CN111461039A (en) * 2020-04-07 2020-07-28 电子科技大学 Landmark identification method based on multi-scale feature fusion

Also Published As

Publication number Publication date
CN111967408A (en) 2020-11-20

Similar Documents

Publication Publication Date Title
CN111539370B (en) Image pedestrian re-identification method and system based on multi-attention joint learning
CN110738207B (en) Character detection method for fusing character area edge information in character image
CN110909673B (en) Pedestrian re-identification method based on natural language description
CN111027493B (en) Pedestrian detection method based on deep learning multi-network soft fusion
Yu et al. High-resolution deep image matting
CN107784288B (en) Iterative positioning type face detection method based on deep neural network
CN106845341B (en) Unlicensed vehicle identification method based on virtual number plate
CN111968150B (en) Weak surveillance video target segmentation method based on full convolution neural network
Bae Object detection based on region decomposition and assembly
CN111178251A (en) Pedestrian attribute identification method and system, storage medium and terminal
CN111967408B (en) Low-resolution pedestrian re-identification method and system based on prediction-recovery-identification
CN112017192B (en) Glandular cell image segmentation method and glandular cell image segmentation system based on improved U-Net network
CN113361432B (en) Video character end-to-end detection and identification method based on deep learning
CN111179270A (en) Image co-segmentation method and device based on attention mechanism
Chen et al. Conditional detr v2: Efficient detection transformer with box queries
CN116311310A (en) Universal form identification method and device combining semantic segmentation and sequence prediction
CN114419406A (en) Image change detection method, training method, device and computer equipment
CN112819011A (en) Method and device for identifying relationships between objects and electronic system
CN112329771A (en) Building material sample identification method based on deep learning
CN114419151A (en) Multi-target tracking method based on contrast learning
CN116258874A (en) SAR recognition database sample gesture expansion method based on depth condition diffusion network
CN116503399B (en) Insulator pollution flashover detection method based on YOLO-AFPS
CN112801029A (en) Multi-task learning method based on attention mechanism
CN112241736A (en) Text detection method and device
CN116168394A (en) Image text recognition method and device

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