Pedestrian snapshot and face snapshot quality evaluation method and device

Abstract

The application discloses a pedestrian snapshot and face snapshot quality evaluation method and device, relates to the picture quality evaluation technical field, and extracts the pedestrian snapshot and the face snapshot with ID by target detection and tracking of the monitoring video, and constructs the pedestrian snapshot data set and the face snapshot data set according to the blur degree and the shielding / missing degree for marking; instance segmentation is carried out based on the snapshot data set, and the instance segmentation mask of the main instance and other instances is determined; the final main instance attention mask is determined according to the instance segmentation mask of the main instance and other instances and the snapshot; the main instance attention snapshot is calculated according to the main instance attention mask and the snapshot, and quality evaluation is carried out according to the main instance attention snapshot, the quality score corresponding to the blur degree and the shielding / missing degree is obtained, the problem that the prior art can only evaluate the face image quality under a single target or simple background is solved, and the stability and reliability of the face recognition system are ensured.

Description

Technical Field

[0001] This application relates to the field of image quality assessment technology, specifically to a method and apparatus for assessing the quality of pedestrian snapshots and face snapshots. Background Technology

[0002] Image quality assessment techniques involve quantifying and predicting human perception of image quality. No-reference image quality assessment techniques (without information about any original reference image or the type of distortion present) aim to make robust and accurate quality predictions that are closely related to subjective judgment. The typical presence of various types of artifacts and the impact of image content on perceived quality make no-reference image quality assessment a challenging problem. It is necessary to be able to objectively determine and control the quality of these digital photographs and guide subsequent processing tasks.

[0003] In recent years, facial image quality assessment has become an indispensable component of facial recognition systems to ensure the stability and reliability of recognition performance in unconstrained scenarios. Under controlled conditions, facial recognition systems typically achieve satisfactory performance. However, in some real-world applications, facial recognition systems need to operate in unconstrained environments (such as surveillance cameras and outdoor scenes), leading to a significant decrease in recognition accuracy and unstable performance. To maintain the stable and reliable performance of facial recognition systems, facial image quality assessment supports the system in selecting high-quality images or deleting low-quality images to stabilize recognition performance.

[0004] Most existing technologies evaluate the quality of face images under a single target or simple background, without considering how to distinguish different targets in complex backgrounds with multiple targets and evaluate only the main target without being affected by other targets and background. Summary of the Invention

[0005] Therefore, this application provides a method and apparatus for evaluating the quality of pedestrian snapshots and face snapshots, in order to solve the problem that existing image quality evaluation techniques can only evaluate the quality of face images with a single target or a simple background.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] Firstly, a method for evaluating the quality of pedestrian and facial snapshots includes:

[0008] Step 1: Perform target detection and tracking on the surveillance video, and extract pedestrian snapshots and face snapshots with IDs;

[0009] Step 2: Annotate the pedestrian snapshots and face snapshots according to their blurriness and occlusion / missingness, and construct pedestrian snapshot datasets and face snapshot datasets;

[0010] Step 3: Perform instance segmentation on the pedestrian snapshot dataset and the face snapshot dataset to obtain the instance segmentation mask for each target;

[0011] Step 4: Determine the instance segmentation masks for the main instance and other instances based on the instance segmentation masks of each target;

[0012] Step 5: Based on the instance segmentation mask and snapshots of the main instance and other instances, identify the relationship between the main instance and other instances. By determining whether there is an occlusion relationship, select other instances that the main instance needs to be fused, and determine the final attention mask of the main instance.

[0013] Step 6: Calculate the main instance attention snapshot based on the main instance attention mask and the snapshot, and perform quality evaluation based on the main instance attention snapshot to obtain the quality scores corresponding to blurriness and occlusion / missingness.

[0014] Preferably, step 1 specifically includes:

[0015] Step 101: Input the surveillance video into the object detection network to regress the bounding boxes of pedestrian and face categories and extract pedestrian snapshots and face snapshots without IDs;

[0016] Step 102: Input the pedestrian snapshots and face snapshots without IDs into the target tracking network and extract the pedestrian snapshots and face snapshots with IDs by combining appearance information through inter-frame prediction and correlation.

[0017] Preferably, in step 101, the target detection network uses YOLOv5.

[0018] Preferably, in step 102, the target tracking network uses StrongSORT.

[0019] Preferably, in step 3, Mask R-CNN is used when segmenting the pedestrian snapshot dataset and the face snapshot dataset into instances.

[0020] Preferably, step 5 specifically includes:

[0021] Step 501: Calculate the intermediate results of the main instance attention of the pedestrian snapshot set and the face snapshot dataset;

[0022] Step 502: Input a snapshot of the intermediate result of the main instance attention into the local semantic segmentation network to obtain the local semantic segmentation mask of the main instance;

[0023] Step 503: Calculate a snapshot of the local segmentation of the intermediate result of the main instance attention and a snapshot of the attention of other instances based on the local semantic segmentation mask of the main instance.

[0024] Step 504: Input the snapshot of the local segmentation of the intermediate result of the main instance attention and the snapshot of the other instance attention into the local relation recognition network to determine whether there is an occlusion relationship between the local semantics and other instances, and output the local relation feature map;

[0025] Step 505: Input the snapshot of the intermediate result of the main instance attention and the snapshot of the other instance attention into the overall relationship recognition network to determine whether there is an occlusion relationship between the main instance and other instances, and output the overall relationship feature map;

[0026] Step 506: Fuse the local relation feature map and the global relation feature map to obtain the occlusion relationship between the main instance and other instances;

[0027] Step 507: Based on the occlusion relationship between the main instance and other instances, merge the main instance with other instances that are not occluded to determine the final main instance attention mask.

[0028] Preferably, the local relation recognition network and the global relation recognition network in steps 504 and 505 both adopt ResNet.

[0029] Secondly, a pedestrian snapshot and facial snapshot quality assessment device includes:

[0030] The detection and tracking unit is used to perform target detection and tracking on the surveillance video, extract pedestrian snapshots and face snapshots with IDs, label the pedestrian snapshots and face snapshots according to blur and occlusion / missing degree, and construct pedestrian snapshot datasets and face snapshot datasets;

[0031] The semantic segmentation unit is used to segment the pedestrian snapshot dataset and the face snapshot dataset into instances, obtain the instance segmentation mask of each target, and determine the instance segmentation mask of the main instance and other instances based on the instance segmentation mask of each target.

[0032] The relationship identification unit is used to identify the relationship between the main instance and other instances based on the instance segmentation mask and snapshot of the main instance and other instances, select other instances that need to be fused by judging whether there is an occlusion relationship, and determine the final main instance attention mask.

[0033] The quality assessment unit is used to calculate the main instance attention snapshot based on the main instance attention mask and the snapshot, and to perform quality assessment based on the main instance attention snapshot to obtain the quality scores corresponding to the blur and occlusion / missing degree.

[0034] Thirdly, a computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a method for assessing the quality of pedestrian snapshots and face snapshots.

[0035] Fourthly, a computer-readable storage medium having a computer program stored thereon, the computer program, when executed by a processor, implementing steps of a method for evaluating the quality of pedestrian snapshots and face snapshots.

[0036] Compared with the prior art, this application has at least the following beneficial effects:

[0037] This application provides a method and apparatus for quality assessment of pedestrian and face snapshots. It extracts ID-bearing pedestrian and face snapshots by performing target detection and tracking on surveillance videos; labels the pedestrian and face snapshots based on ambiguity and occlusion / missing information, and constructs pedestrian and face snapshot datasets; performs instance segmentation on the pedestrian and face snapshot datasets to obtain instance segmentation masks for each target; determines the instance segmentation masks of the main instance and other instances based on the instance segmentation masks of each target; identifies the relationship between the main instance and other instances based on the instance segmentation masks and snapshots, selects other instances to be fused to the main instance by determining whether occlusion relationships exist, and determines the final main instance attention mask; calculates the main instance attention snapshot based on the main instance attention mask and snapshot, and performs quality assessment based on the main instance attention snapshot to obtain quality scores corresponding to ambiguity and occlusion / missing information. This solves the problem that existing technologies can only assess the quality of face images with a single target or simple background, and ensures the stability and reliability of the face recognition system. Attached Figure Description

[0038] To more intuitively illustrate the prior art and this application, several exemplary figures are provided below. It should be understood that the specific shapes and structures shown in the figures should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary figures, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0039] Figure 1 A flowchart illustrating a method for evaluating the quality of pedestrian and facial snapshots provided in Embodiment 1 of this application;

[0040] Figure 2 This is a schematic diagram of the network structure for constructing pedestrian snapshot datasets and face snapshot datasets provided in Embodiment 1 of this application;

[0041] Figure 3A schematic diagram of the network structure for determining the instance segmentation mask of each target as provided in Embodiment 1 of this application;

[0042] Figure 4 A schematic diagram of the network structure for determining the final principal instance attention mask provided in Embodiment 1 of this application;

[0043] Figure 5 This is a schematic diagram of the network structure for determining the quality score provided in Embodiment 1 of this application;

[0044] Figure 6 This is a schematic diagram of a pedestrian snapshot and face snapshot quality assessment device provided in Embodiment 2 of this application. Detailed Implementation

[0045] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "comprising," "including," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0047] The terms used in this application, such as "upper," "lower," "left," "right," and "middle," are generally used to facilitate intuitive understanding by referring to the accompanying drawings, and are not absolute limitations on the positional relationships in the actual product. Changes in these relative positional relationships, without departing from the technical concept disclosed in this application, should also be considered within the scope of this application.

[0048] Example 1

[0049] Please see Figure 1 This embodiment provides a method for evaluating the quality of pedestrian snapshots and face snapshots, including:

[0050] S1: Perform target detection and tracking on the surveillance video, and extract pedestrian snapshots and face snapshots with IDs;

[0051] S101: Input the surveillance video into the object detection network to regress the bounding boxes of pedestrian and face categories to extract pedestrian snapshots and face snapshots without IDs;

[0052] Specifically, YOLOv5 is the preferred object detection network.

[0053] S102: Input pedestrian snapshots and face snapshots without IDs into the target tracking network and extract pedestrian snapshots and face snapshots with IDs by combining appearance information through inter-frame prediction and correlation;

[0054] Specifically, the target tracking network is preferably StrongSORT.

[0055] S2: Label pedestrian snapshots and face snapshots with IDs based on ambiguity and occlusion / missing information, and construct pedestrian snapshot datasets and face snapshot datasets;

[0056] S201: Add blur annotations to pedestrian snapshots and face snapshots with IDs;

[0057] Specifically, when labeling ambiguity, the scoring range is 0 to 5.5 points (the higher the score, the lower the ambiguity), and the average score after multiple people have given their scores is taken.

[0058] S202: Label the occlusion / missingness of pedestrian snapshots and face snapshots with IDs;

[0059] Specifically, when marking occlusion / missing information, the scoring range is 0 to 10 points (the body is divided into 10 parts: head, torso, left upper arm, right upper arm, left forearm, right forearm, left thigh, right thigh, left calf and right calf, each complete part is worth 1 point), and the average score after multiple people give the score is taken.

[0060] S203: Integrate ID-annotated pedestrian snapshots and face snapshots based on ambiguity and occlusion / missing information into a pedestrian snapshot dataset. and face snapshot dataset A detailed network structure diagram is shown below. Figure 2 As shown.

[0061] S3: Perform instance segmentation on the pedestrian snapshot dataset and the face snapshot dataset to obtain the instance segmentation mask for each target;

[0062] Before performing instance segmentation, this step requires calibrating the pedestrian and face snapshots, specifically including:

[0063] S301: Use the feature calibration module to distinguish the contours and occlusion boundaries of instances in the pedestrian snapshot dataset;

[0064] Specifically, for a set of pedestrian snapshot datasets with the same ID

[0065] Will The input feature calibration module is used to distinguish the contours and occlusion boundaries of instances.

[0066] S302: Multiple spatial correlation feature maps are obtained by enumeration, and the spatial correlation feature maps are input into multiple stacked convolutional layers to obtain the spatial calibration offset;

[0067] Specifically, calculate the spatial correlation c(x1,x2) between the patch centered at x1 in frame f1 and the patch centered at x2 in frame f2. o∈[-k,k]×[-k,k] <f1(x1+o),f2(x2+o)>, where the patch size is 2k+1.

[0068] S303: Calibrate the pedestrian snapshot based on the spatial calibration offset to obtain the calibrated pedestrian snapshot;

[0069] Specifically, adding the spatial calibration offset to the original snapshot yields the calibrated snapshot.

[0070] S304: Input the calibrated pedestrian snapshot into the instance segmentation network for instance segmentation to obtain the instance segmentation mask for each target;

[0071] For details, please refer to Figure 3 The calibrated pedestrian snapshots are fed together with the original snapshots into the Mask R-CNN instance segmentation network. Features are extracted from each candidate box through the RoIAlign operation, and classification, bounding box regression, and binary segmentation are performed in parallel using three dedicated branches to obtain the instance segmentation mask.

[0072] S305: Facial snapshots of a group of people with the same ID Repeat S301 to S304.

[0073] S4: Determine the instance segmentation masks for the main instance and other instances based on the instance segmentation masks for each target;

[0074] Specifically, based on the candidate box size corresponding to the instance segmentation mask of each target, the main instance is selected, the candidate box with the largest size is the main instance, and the others are other instances, and instance segmentation masks for the main instance and other instances are generated respectively.

[0075] S5: Based on the instance segmentation mask and snapshots of the main instance and other instances, identify the relationship between the main instance and other instances, select other instances that need to be fused by judging whether there is an occlusion relationship, and determine the final main instance attention mask.

[0076] S501: Pedestrian snapshots: intermediate results of calculating the main instance attention of the pedestrian snapshot set;

[0077] Specifically, for a group of pedestrian snapshots with the same ID Use formula Pedestrian snapshots as intermediate results of calculating primary instance attention Among them, ⊙ represents the Hadamard product.

[0078] S502: Input the pedestrian snapshot, which is the intermediate result of the main instance attention, into the local semantic segmentation network to obtain the local semantic segmentation mask of the main instance;

[0079] Specifically, the pedestrian snapshot is the intermediate result of the main instance attention. Input a local semantic segmentation network. First, a fully convolutional network (FCN) regresses bounding boxes for multiple body parts (head, torso, arms, hands, legs). Then, the bounding boxes are combined with the instance segmentation mask of the main instance to generate a local semantic segmentation mask for the main instance.

[0080] S503: Calculate pedestrian snapshots of the intermediate results of local segmentation of the main instance attention and pedestrian snapshots of other instance attention based on the local semantic segmentation mask of the main instance.

[0081] Specifically, using the formula Local segmentation of pedestrian snapshots for computing intermediate results of primary instance attention

[0082] Use formula Pedestrian snapshots for calculating attention to other instances

[0083] S504: Input the pedestrian snapshots of the local segmentation of the intermediate results of the main instance attention and the pedestrian snapshots of the other instance attention into the local relation recognition network to determine whether there is an occlusion relationship between the local semantics and other instances, and output the local relation feature map.

[0084] Specifically, it involves locally segmenting pedestrian snapshots from the intermediate results of the main instance attention. Pedestrian snapshots and other instance attention Input the ResNet local relation recognition network to determine whether there is an occlusion relationship between the local semantics and other instances, and output the local relation feature map.

[0085] S505: Input the pedestrian snapshot of the intermediate result of the main instance attention and the pedestrian snapshot of the other instance attention into the overall relationship recognition network to determine whether there is an occlusion relationship between the main instance and other instances, and output the overall relationship feature map;

[0086] Specifically, the pedestrian snapshots are the intermediate results of the main instance attention. Pedestrian snapshots and other instance attention Input the ResNet global relationship recognition network to determine whether there is an occlusion relationship between the main instance and other instances, and output the global relationship feature map.

[0087] S506: Fuse the local relation feature map and the global relation feature map to obtain the occlusion relationship between the main instance and other instances;

[0088] Specifically, a support vector machine (SVM) is used to fuse local relation feature maps and global relation feature maps, outputting a fused relation feature map. The occlusion relationship between the main instance and other instances is determined based on the fused relation feature map.

[0089] S507: Based on the occlusion relationship between the main instance and other instances, merge the main instance with other instances that are not occluded to determine the final main instance attention mask;

[0090] Specifically, based on the fused relation feature map, the main instance is fused with other instances with non-occluded relations to determine the final main instance attention mask. A detailed network structure diagram is shown below. Figure 4 As shown.

[0091] S508: Snapshots of a group of faces with the same ID Repeat S501 to S507.

[0092] S6: Calculate the main instance attention snapshot based on the main instance attention mask and the snapshot, and perform quality evaluation based on the main instance attention snapshot to obtain the quality scores corresponding to blurriness and occlusion / missingness.

[0093] S601: Calculate the main instance attention pedestrian snapshot;

[0094] Specifically, for a group of pedestrian snapshots with the same ID Use formula Calculate the primary instance attention pedestrian snapshot Among them, ⊙ represents the Hadamard product.

[0095] S602: Input the main instance attention snapshot into the quality assessment network for quality assessment;

[0096] Specifically, the main instance attention snapshot Input the ResNet quality evaluation network for quality evaluation.

[0097] S603: Network output quality scores for ambiguity and occlusion / missing information;

[0098] Specifically, the network outputs a quality score vector: They represent The corresponding blur and occlusion / missing score.

[0099] S604: Snapshots can be selected based on quality score.

[0100] Specifically, the higher the fuzziness score... (Lower severity) and higher occlusion / missing score (Lower severity) corresponding pedestrian snapshot The higher the image quality, the better the detailed network structure diagram will appear. Figure 5 As shown.

[0101] S605: Facial snapshots of a group of people with the same ID Repeat S601 to S404.

[0102] This embodiment provides a method for evaluating the quality of pedestrian and face snapshots, which solves the problem that existing technologies can only evaluate the quality of face images with a single target or simple background. By segmenting the main instance and other instances and determining whether there is an occlusion relationship, an attention mask for the main instance is generated. This enables the evaluation of only the main target in complex backgrounds with multiple targets, without being affected by other targets and backgrounds, thereby ensuring the stability and reliability of the face recognition system.

[0103] Example 2

[0104] Please see Figure 6 This embodiment provides a pedestrian snapshot and face snapshot quality assessment device, including:

[0105] The detection and tracking unit is used to perform target detection and tracking on the surveillance video, extract pedestrian snapshots and face snapshots with IDs, label the pedestrian snapshots and face snapshots according to blur and occlusion / missing degree, and construct pedestrian snapshot datasets and face snapshot datasets;

[0106] The semantic segmentation unit is used to segment the pedestrian snapshot dataset and the face snapshot dataset into instances, obtain the instance segmentation mask of each target, and determine the instance segmentation mask of the main instance and other instances based on the instance segmentation mask of each target.

[0107] The relationship identification unit is used to identify the relationship between the main instance and other instances based on the instance segmentation mask and snapshot of the main instance and other instances, select other instances that need to be fused by judging whether there is an occlusion relationship, and determine the final main instance attention mask.

[0108] The quality assessment unit is used to calculate the main instance attention snapshot based on the main instance attention mask and the snapshot, and to perform quality assessment based on the main instance attention snapshot to obtain the quality scores corresponding to the blur and occlusion / missing degree.

[0109] For specific limitations regarding a pedestrian snapshot and face snapshot quality assessment device, please refer to the limitations regarding a pedestrian snapshot and face snapshot quality assessment method mentioned above, which will not be repeated here.

[0110] Example 3

[0111] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a method for evaluating the quality of pedestrian snapshots and face snapshots.

[0112] Example 4

[0113] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a method for evaluating the quality of pedestrian snapshots and face snapshots.

[0114] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.

[0115] The present application has been described in a relatively specific and detailed manner above through general descriptions and specific embodiments. It should be understood that, based on the technical concept of the present application, several conventional adjustments or further innovations can be made to these specific embodiments; however, as long as they do not depart from the technical concept of the present application, the technical solutions obtained by these conventional adjustments or further innovations also fall within the protection scope of the claims of the present application.

Claims

1. A method for pedestrian snapshot and face snapshot quality assessment, characterized in that, The application relates to a method for generating a primary instance attention mask of a target in a surveillance video. The method comprises the following steps: Step 1: target detection and tracking are performed on the surveillance video, and IDed pedestrian snapshots and face snapshots are extracted; Step 2: the IDed pedestrian snapshots and face snapshots are labeled according to blur degree and occlusion / loss degree, and a pedestrian snapshot dataset and a face snapshot dataset are constructed; Step 3: instance segmentation is performed on the pedestrian snapshot dataset and the face snapshot dataset to obtain instance segmentation masks of each target; Step 4: the instance segmentation masks of the primary instance and other instances are determined according to the instance segmentation masks of each target; Step 5: the relationship between the primary instance and other instances is identified according to the instance segmentation masks of the primary instance and other instances and the snapshots, the other instances that need to be fused by the primary instance are selected by judging whether there is an occlusion relationship, and the final primary instance attention mask is determined; Step 6: the primary instance attention snapshot is calculated according to the primary instance attention mask and the snapshot, and the quality score corresponding to the blur degree and the occlusion / loss degree is obtained by performing quality evaluation according to the primary instance attention snapshot. The step 2 specifically comprises the following steps: Step 201: the IDed pedestrian snapshots and face snapshots are labeled according to the blur degree; Step 202: the IDed pedestrian snapshots and face snapshots are labeled according to the occlusion / loss degree; 2.The method of claim 1, wherein, Step 203: the IDed pedestrian snapshots and face snapshots labeled according to the blur degree and the occlusion / loss degree are integrated into the pedestrian snapshot dataset and the face snapshot dataset. The step 1 specifically comprises the following steps: Step 101: the surveillance video is input into a target detection network to regress the boundary box of the pedestrian and face category, and IDless pedestrian snapshots and face snapshots are extracted; 3.The method of claim 2, wherein, Step 102: the IDless pedestrian snapshots and face snapshots are input into a target tracking network, and IDed pedestrian snapshots and face snapshots are extracted through inter-frame prediction and association combined with appearance information. 4.The method of claim 2, wherein, In the step 101, the target detection network adopts YOLOv5. 5.The method of claim 1, wherein, In the step 102, the target tracking network adopts StrongSORT.

6. The method of pedestrian snapshot and face snapshot quality assessment according to claim 1, characterized in that, In the step 3, Mask R-CNN is adopted when the pedestrian snapshot dataset and the face snapshot dataset are subjected to instance segmentation. The step 5 specifically comprises the following steps: Step 501: the snapshot of the intermediate result of the primary instance attention of the pedestrian snapshot dataset and the face snapshot dataset is calculated; Step 502: the snapshot of the intermediate result of the primary instance attention is input into a local semantic segmentation network to obtain the local semantic segmentation mask of the primary instance; Step 503: the snapshot of the local segmentation of the intermediate result of the primary instance attention and the snapshot of the other instance attention are calculated according to the local semantic segmentation mask of the primary instance; Step 504: the snapshot of the local segmentation of the intermediate result of the primary instance attention and the snapshot of the other instance attention are input into a local relationship identification network to judge whether there is an occlusion relationship between the local semantics and the other instance, and a local relationship feature map is output; Step 505: the snapshot of the intermediate result of the primary instance attention and the snapshot of the other instance attention are input into an overall relationship identification network to judge whether there is an occlusion relationship between the primary instance and the other instance, and an overall relationship feature map is output. Step 506: fuse the local relationship feature map and the global relationship feature map to obtain the occlusion relationship between the main instance and other instances; Step 507: determine the final main instance attention mask by fusing the non-occlusion relationship other instances of the main instance according to the occlusion relationship between the main instance and other instances.

7. The method of pedestrian snapshot and face snapshot quality assessment according to claim 6, characterized in that, The local relationship identification network and the global relationship identification network in the step 504 and the step 505 both adopt ResNet.

8. A pedestrian snapshot and face snapshot quality evaluation apparatus characterized by comprising: a pedestrian snapshot and face snapshot quality evaluation unit configured to evaluate a quality of a pedestrian snapshot and a face snapshot. Comprise: A detection and tracking unit is configured to detect and track targets in a monitoring video, extract IDed pedestrian snapshots and face snapshots, label the pedestrian snapshots and face snapshots according to blur degree and occlusion / loss degree, and construct a pedestrian snapshot dataset and a face snapshot dataset; wherein the labeling of the pedestrian snapshots and face snapshots according to blur degree and occlusion / loss degree, and the construction of the pedestrian snapshot dataset and the face snapshot dataset specifically include: labeling the IDed pedestrian snapshots and face snapshots according to blur degree; labeling the IDed pedestrian snapshots and face snapshots according to occlusion / loss degree; integrating the IDed pedestrian snapshots and face snapshots labeled according to blur degree and occlusion / loss degree into the pedestrian snapshot dataset and the face snapshot dataset; A semantic segmentation unit is configured to perform instance segmentation on the pedestrian snapshot dataset and the face snapshot dataset to obtain instance segmentation masks of each target, and determine instance segmentation masks of a main instance and other instances according to the instance segmentation masks of each target; A relationship identification unit is configured to identify the relationship between the main instance and other instances according to the instance segmentation masks of the main instance and other instances and snapshots, select other instances that need to be fused by the main instance by judging whether there is an occlusion relationship, and determine a final main instance attention mask; A quality evaluation unit is configured to calculate a main instance attention snapshot according to the main instance attention mask and snapshots, and perform quality evaluation according to the main instance attention snapshot to obtain quality scores corresponding to blur degree and occlusion / loss degree. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7. The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7.

Images