Picture screening method and device

Abstract

The application provides a picture screening method and device, relates to the technical field of picture processing, and obtains a picture library to be screened; a target picture to be screened is input into a first target detection model to obtain a first feature activation map set; the target picture to be screened is input into a second target detection model to obtain a second feature activation map set, and the accuracy of the second target detection model is higher than that of the first target detection model; a feature value of a first feature activation map in the first feature activation map set and a first weight are calculated to determine a first attention degree of the first feature activation map set; similarly, a second attention degree of the second feature activation map set is determined; a complexity of the target picture to be screened is calculated according to a difference between the first attention degree and the second attention degree; if the complexity exceeds a preset complexity threshold, the target picture to be screened is retained; otherwise, the target picture to be screened is deleted. In this way, automatic screening of pictures can be realized, and the picture screening processing efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of picture processing, in particular to a picture screening method and device. BACKGROUND

[0002] When analyzing pictures in a vehicle driving scene, the picture processing takes a long time due to a large number of pictures, and there are a large number of simple pictures with simple content and no analysis value in the pictures, resulting in low efficiency of obtaining effective information from the pictures.

[0003] Currently, pictures in a vehicle driving scene are manually screened first to exclude simple pictures, but manual screening is time-consuming and laborious, and the labor cost is high. SUMMARY

[0004] Therefore, the embodiments of the present application provide a picture screening method and device, aiming to improve the picture screening efficiency.

[0005] In a first aspect, the embodiments of the present application provide a picture screening method, which comprises:

[0006] obtaining a to-be-screened picture library, wherein the to-be-screened picture library includes at least one to-be-screened picture;

[0007] inputting a target to-be-screened picture into a first target detection model to obtain a first feature activation map set including a plurality of first feature activation maps through the first target detection model;

[0008] inputting the target to-be-screened picture into a second target detection model to obtain a second feature activation map set including a plurality of second feature activation maps through the second target detection model, wherein the accuracy of the second target detection model is higher than that of the first target detection model;

[0009] calculating a first weight of the first feature activation map in the first feature activation map set and a second weight of each second feature activation map in the second feature activation map set;

[0010] determining a first attention degree of the first feature activation map set according to the feature value of the first feature activation map in the first feature activation map set and the first weight;

[0011] determining a second attention degree of the second feature activation map set according to the feature value of the second feature activation map in the second feature activation map set and the second weight;

[0012] if the difference between the first attention degree and the second attention degree exceeds a preset complexity threshold, retaining the target to-be-screened picture; and if the difference between the first attention degree and the second attention degree does not exceed the preset complexity threshold, deleting the target to-be-screened picture from the to-be-screened picture library.

[0013] Optionally, the first target detection model and the second target detection model are the same type of target detection model.

[0014] Optionally, the first target detection model is trained with a smaller amount of data than the second target detection model, and / or the first target detection model is trained for a smaller number of times than the second target detection model.

[0015] Optionally, the first weight of the first feature activation map in the first feature activation map set is calculated by:

[0016] The first feature activation map set is solved by singular value decomposition (SVD) to obtain a right singular value matrix;

[0017] The numerical value of the first row of the right singular value matrix is taken as the first weight of each first feature activation map in the first feature activation map set;

[0018] Or,

[0019] The target to-be-screened picture is input into the first target detection model to obtain first initial data of the first target detection model output identifying the to-be-identified object in the target to-be-screened picture;

[0020] The eigenvalue of the target first feature activation map in the first feature activation map set is set to zero or a minimum value to obtain a perturbed first feature activation map set;

[0021] The perturbed first feature activation map set is input into the data decision layer of the first target detection model to obtain first perturbed data output by the data decision layer of the first target detection model;

[0022] According to the first perturbed data and the first initial data, deviation data of the to-be-identified object in the target to-be-screened picture is obtained;

[0023] According to the deviation data, the first weight of the target first feature activation map is determined;

[0024] According to the eigenvalue of the first feature activation map in the first feature activation map set and the first weight, the first attention degree of the first feature activation map set is determined, including:

[0025] The eigenvalue of each first feature activation map in the first feature activation map set is multiplied by the first weight of the first feature activation map to obtain a feature weight value of the first feature activation map;

[0026] The feature weight values of the first feature activation maps included in the first feature activation map set are summed to obtain the first attention degree of the first feature activation map set.

[0027] Optionally, the calculating the second weight of the second feature activation map in the second feature activation map set comprises:

[0028] The second feature activation map set is solved by singular value decomposition (SVD) to obtain a right singular value matrix;

[0029] The numerical value of the first row of the right singular value matrix is taken as the second weight of each second feature activation map in the second feature activation map set;

[0030] Or,

[0031] The target picture to be screened is input into a second target detection model to obtain second initial data of the second target detection model output identifying the object to be identified in the target picture to be screened;

[0032] The eigenvalue of the target second feature activation map in the second feature activation map set is set to zero or a minimum value to obtain a perturbed second feature activation map set;

[0033] The perturbed second feature activation map set is input into the data decision layer of the second target detection model to obtain second perturbed data output by the data decision layer of the second target detection model;

[0034] According to the second perturbed data and the second initial data, deviation data of the object to be identified in the target picture to be screened is obtained;

[0035] According to the deviation data, the second weight of the target second feature activation map is determined;

[0036] The second attention degree of the second feature activation map set is determined according to the eigenvalue of the second feature activation map in the second feature activation map set and the second weight, comprising:

[0037] The eigenvalue of each second feature activation map in the second feature activation map set is multiplied by the second weight of the second feature activation map to obtain a feature weight value of the second feature activation map;

[0038] The feature weight values of the second feature activation maps included in the second feature activation map set are summed to obtain the second attention degree of the second feature activation map set.

[0039] Optionally, the calculating the first weight of the first feature activation map in the first feature activation map set comprises:

[0040] The first number of the first feature activation maps included in the first feature activation map set is obtained;

[0041] If the first number exceeds a preset number threshold,

[0042] obtaining a right singular value matrix by solving the first feature activation map set using singular value decomposition (SVD);

[0043] using the numerical values of the first row of the right singular value matrix as the weights of each first feature activation map in the first feature activation map set;

[0044] sorting the weights of all first feature activation maps in the first feature activation map set, and selecting a preset number of first feature activation maps in the front of the sorting to form a first target activation map set;

[0045] inputting the target picture to be screened into the first target detection model to obtain first initial data of identifying the object to be identified in the target picture to be screened output by the first target detection model;

[0046] setting the eigenvalues of the target first feature activation map in the first target activation map set to zero or a minimum value to obtain a perturbed first target activation map set;

[0047] inputting the perturbed first target activation map set into the data decision layer of the first target detection model to obtain first perturbed data output by the data decision layer of the first target detection model;

[0048] obtaining deviation data of the object to be identified in the target picture to be screened according to the first perturbed data and the first initial data;

[0049] determining a first weight of the target first feature activation map according to the deviation data;

[0050] if the first number does not exceed a preset number threshold,

[0051] inputting the target picture to be screened into the first target detection model to obtain first initial data of identifying the object to be identified in the target picture to be screened output by the first target detection model;

[0052] setting the eigenvalues of the target first feature activation map in the first feature activation map set to zero or a minimum value to obtain a perturbed first feature activation map set;

[0053] inputting the perturbed first feature activation map set into the data decision layer of the first target detection model to obtain first perturbed data output by the data decision layer of the first target detection model;

[0054] obtaining deviation data of the object to be identified in the target picture to be screened according to the first perturbed data and the first initial data;

[0055] determining a first weight of the target first feature activation map according to the deviation data.

[0056] Optionally, the first attention degree of the first feature activation map set is determined according to the feature values of the first feature activation map in the first feature activation map set and the first weight, and the determining comprises:

[0057] The feature weight value of each first feature activation map to be analyzed in the first target activation map set is obtained by multiplying the feature value of the first feature activation map to be analyzed by the first weight of the first feature activation map to be analyzed.

[0058] The first attention degree of the first target activation map set is obtained by summing the feature weight values of the first feature activation maps to be analyzed included in the first target activation map set, and the first attention degree of the first target activation map set is taken as the first attention degree of the first feature activation map set.

[0059] Optionally, the second weight of the second feature activation map in the second feature activation map set is calculated, and the calculating comprises:

[0060] The second number of the second feature activation maps included in the second feature activation map set is obtained.

[0061] If the second number exceeds a preset number threshold,

[0062] The second feature activation map set is solved by singular value decomposition (SVD) to obtain a right singular value matrix.

[0063] The first row of the right singular value matrix is taken as the weight of each second feature activation map in the second feature activation map set.

[0064] The weights of all second feature activation maps in the second feature activation map set are sorted, and a set number of second feature activation maps in the front of the sorting are selected to form a second target activation map set.

[0065] The target to-be-screened picture is input into the second target detection model to obtain second initial data of the second target detection model output identifying the to-be-identified object in the target to-be-screened picture.

[0066] The feature values of the target second feature activation map in the second target activation map set are set to zero or a minimum value to obtain a perturbed second target activation map set.

[0067] The perturbed second target activation map set is input into the data decision layer of the second target detection model to obtain second perturbation data output by the data decision layer of the second target detection model.

[0068] According to the second perturbation data and the second initial data, deviation data of the to-be-identified object in the target to-be-screened picture is obtained.

[0069] According to the deviation data, the second weight of the target second feature activation map is determined.

[0070] if the second quantity does not exceed a preset quantity threshold,

[0071] input the target to-be-screened picture into a second target detection model to obtain second initial data of recognizing an object to be recognized in the target to-be-screened picture output by the second target detection model;

[0072] set a feature value of a target second feature activation map in the second feature activation map set to zero or a minimum value to obtain a perturbed second feature activation map set;

[0073] input the perturbed second feature activation map set into a data decision layer of the second target detection model to obtain second perturbed data output by the data decision layer of the second target detection model;

[0074] obtain deviation data of the object to be recognized in the target to-be-screened picture according to the second perturbed data and the second initial data;

[0075] determine a second weight of the target second feature activation map according to the deviation data.

[0076] In a second aspect, the present application further provides a picture screening device, comprising:

[0077] an acquisition module configured to acquire a to-be-screened picture library, the to-be-screened picture library comprising at least one to-be-screened picture;

[0078] a first processing module configured to input a target to-be-screened picture into a first target detection model to obtain a first feature activation map set comprising a plurality of first feature activation maps through the first target detection model;

[0079] a second processing module configured to input the target to-be-screened picture into a second target detection model to obtain a second feature activation map set comprising a plurality of second feature activation maps through the second target detection model, the accuracy of the second target detection model being higher than that of the first target detection model;

[0080] a first calculation module configured to calculate a first weight of a first feature activation map in the first feature activation map set and a second weight of each second feature activation map in the second feature activation map set;

[0081] a second calculation module configured to determine a first attention degree of the first feature activation map set according to a feature value of the first feature activation map in the first feature activation map set and the first weight;

[0082] a third calculation module configured to determine a second attention degree of the second feature activation map set according to a feature value of the second feature activation map in the second feature activation map set and the second weight;

[0083] The judgment module is configured to: if the difference between the first attention degree and the second attention degree exceeds a preset complexity threshold, retaining the target picture to be screened; and if the difference between the first attention degree and the second attention degree does not exceed the preset complexity threshold, deleting the target picture to be screened from the picture library to be screened.

[0084] The embodiment of the present application provides a picture screening method and device. The embodiment of the present application acquires a picture library to be screened, inputs a target picture to be screened into a first target detection model and a second target detection model which have different capabilities and different model result accuracy respectively, analyzes the same target picture to be screened by using two models with different capability levels, and acquires a first feature activation map set output by the first target detection model and a second feature activation map set output by the second target detection model. The first attention degree of the first feature activation map set is determined according to the feature value of the first feature activation map in the first feature activation map set and the first weight, and the second attention degree of the second feature activation map set is determined in the same way. The complexity of the target picture to be screened is determined by calculating the difference between the first attention degree and the second attention degree. If the complexity exceeds a preset complexity threshold, the target picture to be screened is retained. If the complexity does not exceed the preset complexity threshold, the target picture to be screened is deleted. The complexity of the target picture to be screened can be effectively analyzed by using the two target detection models with different capability levels. The picture with high complexity is retained, and the picture with low complexity is deleted. The automatic screening of the picture is realized, and the picture screening processing efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0085] To make the technical solutions in the embodiments or the prior art clearer, the accompanying drawings needed in the embodiments or the prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only show some embodiments of the present application, and other accompanying drawings can be obtained by those skilled in the art without any creative effort.

[0086] Figure 1 A picture screening method flowchart provided by the embodiment of the present application;

[0087] Figure 2 A structure diagram of a picture screening device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0088] When pictures in a vehicle driving scene are analyzed, the picture processing time is long due to a large number of pictures, and a large number of simple pictures with simple content and without analysis value exist in the pictures, so that the efficiency of obtaining effective information from the pictures is low. At present, the pictures in the vehicle driving scene are manually screened first, and the simple pictures are removed. However, the manual screening is time-consuming and laborious, and the labor cost is high.

[0089] The application processes the same to-be-screened picture by using the high and low two levels of target detection models. When analyzing a simple picture, the difference in attention obtained by the high and low levels of target detection models is small. When analyzing a complex picture, the difference between the two feature activation maps obtained by the two levels of target detection models is large. Then, the first attention and the second attention are analyzed according to the feature activation maps output by the two levels of target detection models. The difference between the first attention and the second attention is large, and the complexity of the to-be-screened picture is high. In this way, the to-be-screened picture can be determined to be retained or deleted according to the complexity, so as to realize the screening of the picture without manual processing, and improve the screening efficiency.

[0090] Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0091] Figure 1 A flowchart of a picture screening method provided by an embodiment of the present application is shown in the figure. The method includes the following steps.

[0092] S101, acquire a to-be-screened picture library, wherein the to-be-screened picture library includes at least one to-be-screened picture.

[0093] The to-be-screened picture can be a picture captured in a real scene, a picture related to an analysis scene downloaded from the Internet, or a picture obtained from other picture acquisition channels.

[0094] S102, input a target to-be-screened picture into a first target detection model, and obtain a first feature activation map set including a plurality of first feature activation maps through the first target detection model.

[0095] The type of the first target detection model can be a region-based convolutional neural network (faster-rcnn), a target detection algorithm such as yoloV7, or other types of target detection models.

[0096] S103, input the target to-be-screened picture into a second target detection model, and obtain a second feature activation map set including a plurality of second feature activation maps through the second target detection model, wherein the accuracy of the second target detection model is higher than that of the first target detection model.

[0097] The type of the second target detection model is the same as that of the first target detection model.

[0098] The accuracy of the first target detection model is higher than the accuracy of the second target detection model can be that the difference between the accuracy of the first target detection model and the accuracy of the second target detection model exceeds a set threshold.

[0099] The accuracy of the first target detection model and the accuracy of the second model are different, in one possible implementation, the different training can be performed on the number of model training times, in another possible implementation, the different training can also be performed on the data amount of model training.

[0100] S104, calculate the first weight of the first feature activation map in the first feature activation map set and the second weight of the second feature activation map in the second feature activation map set.

[0101] The first weight and the second weight can be calculated by using the method of perturbation (for example, the feature value of the target first feature activation map in the first feature activation map set can be adjusted, the recognition result output by the target detection model is obtained, and the result output by the target detection model is compared with the to-be-screened picture, the first weight of the target first feature activation map is determined according to the deviation of the two results, and the second weight of the second feature activation map is obtained by the same method as the first weight), the first weight and the second weight can also be calculated by using the method of non-perturbation (for example, singular value decomposition can be used), and of course the first weight and the second weight can also be calculated by using the method of combining perturbation and non-perturbation.

[0102] S105, according to the feature value of the first feature activation map in the first feature activation map set and the first weight, determine the first attention degree of the first feature activation map set.

[0103] In one possible implementation, the feature value of each first feature activation map in the first feature activation map set used for calculating the first attention degree is multiplied by the weight of the first feature activation map, to obtain the feature weight value of the first feature activation map, and then the feature weight values of all the first feature activation maps in the first feature activation map set used for calculating the first attention degree are summed, to obtain the first attention degree of the first feature activation map.

[0104] S106, according to the feature value of the second feature activation map in the second feature activation map set and the second weight, determine the second attention degree of the second feature activation map set.

[0105] In one possible implementation, the feature value of each second feature activation map in the second feature activation map set used for calculating the second attention degree is multiplied by the weight of the second feature activation map, to obtain the feature weight value of the second feature activation map, and then the feature weight values of all the second feature activation maps in the second feature activation map set used for calculating the second attention degree are summed, to obtain the second attention degree of the second feature activation map.

[0106] S107. If the difference between the first level of attention and the second level of attention exceeds a preset complexity threshold, the target image to be filtered is retained; if the difference between the first level of attention and the second level of attention does not exceed the preset complexity threshold, the target image to be filtered is deleted from the image library to be filtered.

[0107] As described in steps S101-S107 above, using two target detection models with different capabilities to process the same image to be screened, when analyzing simple images, the difference in attention gained by the two models is minimal. However, when analyzing complex images, the two feature activation maps obtained by the two models differ significantly. Therefore, based on the feature activation maps output by the two models, the difference between the first and second attention levels is substantial, indicating a high complexity of the image to be screened. This allows for the determination of whether to retain or delete images based on their complexity, thus achieving image screening without manual processing and improving screening efficiency.

[0108] In the embodiments of this application, the above Figure 1 In steps S104-S106, the attention level of the images to be filtered is calculated. There are possible implementation methods, which are described below. It should be noted that the implementation methods given below are merely illustrative examples and do not represent all implementation methods of the embodiments of this application.

[0109] In one possible implementation, the determination of the first and second attention scores, as well as the determination of the weights of the feature activation maps in the first and second feature activation map sets, can both be done in a non-perturbative manner, or by perturbing the feature values ​​of the target feature activation maps in the feature activation map sets to calculate the weights of the target feature activation maps. The specific methods for obtaining the first and second attention scores of the images to be screened are as follows:

[0110] A1. Calculate the first weight of the first feature activation map in the first feature activation map set, including:

[0111] The first feature activation map set is solved by Singular Value Decomposition (SVD) to obtain the right singular value matrix; the values ​​of the first row of the right singular value matrix are used as the first weights of each first feature activation map in the first feature activation map set.

[0112] or,

[0113] The target image to be screened is input into a first target detection model to obtain the first initial data output by the first target detection model for identifying objects to be identified in the target image to be screened; the feature values ​​of the target first feature activation map in the first feature activation map set are set to zero or a minimum value to obtain a perturbed first feature activation map set; the perturbed first feature activation map set is input into the data decision layer of the first target detection model to obtain the first perturbed data output by the data decision layer of the first target detection model; based on the first perturbed data and the first initial data, the deviation data of the objects to be identified in the target image to be screened is obtained; and the first weight of the target first feature activation map is determined based on the deviation data.

[0114] In one possible implementation, the first initial data includes bounding boxes and recognition results of objects to be identified in the image to be screened, identified by a first object detection model. The first perturbation data includes bounding boxes and recognition results of objects to be identified by the first object detection model based on a first feature activation atlas of the perturbation. The objects to be identified can be vehicles, pedestrians, or other pre-labeled objects trained in the object detection model.

[0115] Based on the second perturbation data and the second initial data, deviation data of the objects to be identified in the target image to be screened is obtained. Specifically, the deviation data includes the offset rate M of the bounding box of the object to be identified by the first target detection model (e.g., IOU) and the probability distribution deviation N of the recognition result (e.g., cross-entropy). The first weight of the target's first feature activation map is determined based on the deviation data. Specifically, the first weight of the target's first feature activation map is obtained by multiplying the bounding box offset rate M by a first coefficient, and then adding the sum of the products of the probability distribution deviation N of the recognition result and a second coefficient. The first and second coefficients are set according to human needs.

[0116] According to step A1 above, in one possible implementation, the weights of the first feature activation map can be obtained using a non-perturbative method, such as Singular Value Decomposition (SVD), to determine the first weights of the first feature activation map. In another possible implementation, the weights of the first feature activation map can also be obtained using a perturbative method. The image to be screened obtains first initial data through a first object detection model. The perturbed first feature activation map set (the perturbed first feature activation map set refers to the set where the feature values ​​of the target's first feature activation map have been adjusted to zero or minimum values) is passed through the data decision layer of the first object detection model to obtain first perturbed data. Finally, the deviation value between the first initial data and the first perturbed data is calculated to obtain the first weight of the target's first feature activation map.

[0117] A2. Calculate the second weights of the second feature activation maps in the second feature activation map set, including:

[0118] The second feature activation map set is solved using Singular Value Decomposition (SVD) to obtain the right singular value matrix; the values ​​in the first row of the right singular value matrix are used as the second weights of each second feature activation map in the second feature activation map set.

[0119] or,

[0120] The target image to be screened is input into a second target detection model to obtain second initial data for identifying objects in the target image to be screened, output by the second target detection model. The feature values ​​of the target second feature activation map in the second feature activation map set are set to zero or a minimum value to obtain a perturbed second feature activation map set. The perturbed second feature activation map set is input into the data decision layer of the second target detection model to obtain second perturbed data output by the data decision layer of the second target detection model. Based on the second perturbed data and the second initial data, deviation data of objects to be identified in the target image to be screened is obtained. The second weight of the target second feature activation map is determined based on the deviation data.

[0121] In one possible implementation, the second initial data includes bounding boxes and recognition results of objects to be identified in the image to be screened, identified by a second object detection model. The second perturbation data includes bounding boxes and recognition results of objects to be identified by the second object detection model based on a second feature activation atlas of the perturbation. The objects to be identified can be vehicles, pedestrians, or other pre-labeled objects trained in the object detection model.

[0122] Based on the second perturbation data and the second initial data, deviation data of the objects to be identified in the target image to be screened is obtained. Specifically, the deviation data includes the offset rate X of the bounding box of the object to be identified by the second target detection model (e.g., IOU) and the probability distribution deviation Y of the recognition result (e.g., cross-entropy). The second weight of the target's second feature activation map is determined based on the deviation data. Specifically, the second weight of the target's second feature activation map is obtained by multiplying the bounding box offset rate X by a first coefficient, and then adding the sum of the products of the probability distribution deviation Y of the recognition result and a second coefficient. The first and second coefficients are set according to human needs.

[0123] The method for obtaining the weights of the second feature activation map in step A2 is the same as the method for obtaining the weights of the first feature activation map in step A1. In one possible implementation, a non-perturbation method, such as singular value decomposition (SVD), can be used. In another possible implementation, a perturbation method can also be used to determine the second weights of the target second feature activation map.

[0124] A3. Based on the feature values ​​and first weights of the first feature activation maps in the first feature activation map set, determine the first level of attention for the first feature activation map set, including:

[0125] Multiply the feature value of each first feature activation map in the first feature activation map set by the first weight of the first feature activation map to obtain the feature weight value of the first feature activation map; sum the feature weight values ​​of the first feature activation maps included in the first feature activation map set to obtain the first attention of the first feature activation map set.

[0126] For example, if the first feature activation map set includes first feature activation map A, first feature activation map B, first feature activation map C, and first feature activation map D, then the first attention calculation process for the first feature activation map set is as follows: first, calculate the first feature weight value obtained by multiplying the feature value of first feature activation map A by the first weight of first feature activation map A; second, calculate the second feature weight value obtained by multiplying the feature value of first feature activation map B by the first weight of first feature activation map B; third, calculate the third feature weight value obtained by multiplying the feature value of first feature activation map C by the first weight of first feature activation map C; and fourth, calculate the fourth feature weight value obtained by multiplying the feature value of first feature activation map D by the first weight of first feature activation map D. Then, sum the first feature weight value, second feature weight value, fourth feature weight value, and fourth feature weight value.

[0127] A4. Based on the feature values ​​and second weights of the second feature activation maps in the second feature activation map set, determine the second level of attention for the second feature activation map set, including:

[0128] Multiply the feature value of each second feature activation map in the second feature activation map set by the second weight of the second feature activation map to obtain the feature weight value of the second feature activation map; sum the feature weight values ​​of the second feature activation maps included in the second feature activation map set to obtain the second attention of the second feature activation map set.

[0129] For example, if the second feature activation map set includes second feature activation map E, second feature activation map F, second feature activation map G, and second feature activation map H, then the second attention calculation process for the second feature activation map set is as follows: first, calculate the fifth feature weight value obtained by multiplying the feature value of second feature activation map A by the second weight of second feature activation map A; the sixth feature weight value obtained by multiplying the feature value of second feature activation map B by the second weight of second feature activation map B; the seventh feature weight value obtained by multiplying the feature value of second feature activation map C by the second weight of second feature activation map C; and the eighth feature weight value obtained by multiplying the feature value of second feature activation map D by the second weight of second feature activation map D. Then, sum the fifth feature weight value, the sixth feature weight value, the seventh feature weight value, and the eighth feature weight value.

[0130] According to steps A1-A4 above, the first weight of the first feature activation map of the first feature activation map set and the second weight of the second feature activation map of the second feature activation map set are obtained first by perturbation or non-perturbation method, and then the first attention of the first feature activation map set and the second attention of the second feature activation map set are obtained.

[0131] In another possible implementation, based on the number of feature maps in the feature map set, the weights of the feature activation maps in both the first and second feature activation map sets can be determined using a non-perturbative method to select the feature activation maps with higher weights to form a target feature map set. Then, based on the target feature map set, the weights of the target feature activation maps are calculated by perturbing and adjusting the feature values ​​of the target feature activation maps in the target feature map set. Specifically, the method for obtaining the first and second attention scores of the image to be selected is as follows:

[0132] B1. Calculate the first weight of the first feature activation map in the first feature activation map set, including:

[0133] Obtain the first number of first feature activation maps included in the first feature activation map set;

[0134] If the first quantity exceeds a preset threshold, Singular Value Decomposition (SVD) is used to solve the first feature activation map set to obtain a right singular value matrix. The values ​​in the first row of the right singular value matrix are used as the weights of each first feature activation map in the first feature activation map set. The weights of all the first feature activation maps in the first feature activation map set are sorted, and a set number of first feature activation maps at the top of the sorting are selected to form a first target activation map set. The target image to be screened is input into a first target detection model to obtain the first initial data output by the first target detection model for identifying the objects to be identified in the target image to be screened. The feature values ​​of the target first feature activation maps in the first target activation map set are set to zero or a minimum value to obtain a perturbed first target activation map set. The perturbed first target activation map set is input into the data decision layer of the first target detection model to obtain the first perturbed data output by the data decision layer of the first target detection model. Based on the first perturbed data and the first initial data, the deviation data of the objects to be identified in the target image to be screened is obtained. The first weight of the target first feature activation map is determined based on the deviation data.

[0135] If the first quantity does not exceed a preset quantity threshold, the target image to be screened is input into the first target detection model to obtain the first initial data output by the first target detection model to identify the object to be identified in the target image to be screened; the feature value of the target first feature activation map in the first feature activation map set is set to zero or a minimum value to obtain a perturbed first feature activation map set; the perturbed first feature activation map set is input into the data decision layer of the first target detection model to obtain the first perturbed data output by the data decision layer of the first target detection model; based on the first perturbed data and the first initial data, the deviation data of the object to be identified in the target image to be screened is obtained; the first weight of the target first feature activation map is determined based on the deviation data.

[0136] In one possible implementation, the first initial data includes bounding boxes and recognition results of objects to be identified in the image to be screened, identified by a first object detection model. The first perturbation data includes bounding boxes and recognition results of objects to be identified by the first object detection model based on a perturbed first object activation atlas. The objects to be identified can be vehicles, pedestrians, or other pre-labeled objects trained in the object detection model.

[0137] Based on the second perturbation data and the second initial data, deviation data of the objects to be identified in the target image to be screened is obtained. Specifically, the deviation data includes the offset rate M of the bounding box of the object to be identified by the first target detection model (e.g., IOU) and the probability distribution deviation N of the recognition result (e.g., cross-entropy). The first weight of the target's first feature activation map is determined based on the deviation data. Specifically, the first weight of the target's first feature activation map is obtained by multiplying the bounding box offset rate M by a first coefficient, and then adding the sum of the products of the probability distribution deviation N of the recognition result and a second coefficient. The first and second coefficients are set according to human needs.

[0138] According to step B1 above, in calculating the first weight of the first feature activation map in the first feature activation map set, it can be first determined whether the number of first feature activation maps included in the first feature activation map set exceeds a preset threshold. If it does, the computational workload of obtaining the first weight through the perturbation method is large. Therefore, a non-perturbation method, such as singular value decomposition (SVD), can be used first to select the first feature activation maps with higher weight values ​​to form the first target activation map set. Then, the first initial data is obtained by the first target detection model based on the image to be selected. The first perturbation data is obtained by the first target detection model and the perturbed first feature activation map set (the perturbed first feature activation map set refers to the map set in which the feature values ​​of the target first feature activation maps in the first feature activation map set have been adjusted to zero or minimum values). The deviation value of the first initial data and the first perturbed data is calculated to obtain the first weight of the target first feature activation map. This improves the computational efficiency while ensuring the accuracy of the first weight calculation. However, if the first number does not exceed the preset number threshold time, then each first feature activation map in the first feature activation map set can be used as the target first feature activation map, and the feature values ​​can be adjusted to perturb it to obtain the first weight of each first feature activation map in the first feature activation map set. By judging whether the first number of first feature activation maps included in the first feature activation map set exceeds the preset number threshold, if the first number exceeds the number threshold, then the first feature activation map with the higher weight is selected for perturbation. If the first number does not exceed the number threshold time, then all first feature activation maps in the first feature activation map set are fully perturbed to achieve a balance between computational efficiency and weight accuracy.

[0139] B2. Calculate the second weights of the second feature activation maps in the second feature activation map set, including:

[0140] Obtain the second number of second feature activation maps included in the second feature activation map set;

[0141] If the second quantity exceeds the preset quantity threshold...

[0142] The second feature activation map set is solved using Singular Value Decomposition (SVD) to obtain a right singular value matrix. The values ​​in the first row of the right singular value matrix are used as the weights of each second feature activation map in the second feature activation map set. The weights of all the second feature activation maps in the second feature activation map set are sorted, and a predetermined number of the top-ranked second feature activation maps are selected to form a second target activation map set. The target image to be screened is input into a second target detection model to obtain second initial data for identifying objects in the target image to be screened. The feature values ​​of the target second feature activation maps in the second target activation map set are set to zero or minimum values ​​to obtain a perturbed second target activation map set. The perturbed second target activation map set is input into the data decision layer of the second target detection model to obtain second perturbation data output by the data decision layer of the second target detection model. Based on the second perturbation data and the second initial data, the deviation data of the objects to be identified in the target image to be screened is obtained. The second weight of the target second feature activation map is determined based on the deviation data.

[0143] If the second quantity does not exceed the preset quantity threshold

[0144] The target image to be screened is input into a second target detection model to obtain second initial data for identifying objects in the target image to be screened, output by the second target detection model. The feature values ​​of the target second feature activation map in the second feature activation map set are set to zero or a minimum value to obtain a perturbed second feature activation map set. The perturbed second feature activation map set is input into the data decision layer of the second target detection model to obtain second perturbed data output by the data decision layer of the second target detection model. Based on the second perturbed data and the second initial data, deviation data of objects to be identified in the target image to be screened is obtained. The second weight of the target second feature activation map is determined based on the deviation data.

[0145] The method for calculating the second weight of the second feature activation map in the second feature activation map set in step B2 is the same as the method for calculating the first weight of the first feature activation map in the first feature activation map set in step B1. By judging whether the second number of the second feature activation map included in the second feature activation map set exceeds the preset number threshold, if the second number exceeds the number threshold, the second feature activation map with high weight is selected for perturbation. If the second number does not exceed the number threshold, full perturbation calculation is performed on all the second feature activation maps in the second feature activation map set to achieve a balance between computational efficiency and weight accuracy.

[0146] In one possible implementation, the second initial data includes bounding boxes and recognition results of objects to be identified in the image to be screened, identified by a second object detection model. The second perturbation data includes bounding boxes and recognition results of objects to be identified by the second object detection model based on a perturbed second object activation atlas.

[0147] Based on the second perturbation data and the second initial data, deviation data of the objects to be identified in the target image to be screened is obtained. Specifically, the deviation data includes the offset rate X of the bounding box of the object to be identified by the second target detection model (e.g., IOU) and the probability distribution deviation Y of the recognition result (e.g., cross-entropy). The second weight of the target's second feature activation map is determined based on the deviation data. Specifically, the second weight of the target's second feature activation map is obtained by multiplying the bounding box offset rate X by a first coefficient, and then adding the sum of the products of the probability distribution deviation Y of the recognition result and a second coefficient. The first and second coefficients are set according to human needs.

[0148] B3. Determine the first attention level of the first feature activation map set based on the feature values ​​and first weights of the first feature activation map set, including:

[0149] Multiply the feature value of each first feature activation map to be analyzed in the first target activation map set by the first weight of the first feature activation map to be analyzed to obtain the feature weight value of the first feature activation map to be analyzed; sum the feature weight values ​​of the first feature activation maps to be analyzed included in the first target activation map set to obtain the first attention of the first target activation map set; and use the first attention of the first target activation map set as the first attention of the first feature activation map set.

[0150] For example, the specific calculation method for the first attention of the first target activation map set as the first attention of the first feature activation map set is as follows: calculate the feature weight value of the first feature activation map A to be analyzed (obtained by multiplying the feature value of the first feature activation map A to be analyzed by the first weight of the first feature activation map A to be analyzed) and the feature weight value of the first feature activation map B to be analyzed (obtained by multiplying the feature value of the first feature activation map B to be analyzed by the first weight of the first feature activation map B to be analyzed), and then sum the feature weight value of the first feature activation map A to be analyzed and the feature weight value of the first feature activation map B to be analyzed.

[0151] B4. The step of determining the second attention level of the second feature activation map set based on the feature values ​​and second weights of the second feature activation map set includes:

[0152] Multiply the feature value of each second feature activation map to be analyzed in the second target activation map set by the second weight of the second feature activation map to be analyzed to obtain the feature weight value of the second feature activation map to be analyzed; sum the feature weight values ​​of the second feature activation maps to be analyzed included in the second target activation map set to obtain the second attention of the second target activation map set; and use the second attention of the second target activation map set as the second attention of the second feature activation map set.

[0153] For example, the specific calculation method for the second attention of the second target activation map set as the second attention of the second feature activation map set is as follows: calculate the feature weight value of the second feature activation map E to be analyzed (obtained by multiplying the feature value of the second feature activation map E to be analyzed by the second weight of the second feature activation map E to be analyzed) and the feature weight value of the second feature activation map F to be analyzed (obtained by multiplying the feature value of the second feature activation map F to be analyzed by the second weight of the second feature activation map F to be analyzed), and then sum the feature weight value of the second feature activation map E to be analyzed and the feature weight value of the second feature activation map F to be analyzed.

[0154] As described in steps B1-B4 above, a first target activation map set, formed by the first feature activation map with a higher weight, and a second target activation map set, formed by the second feature activation map with a higher weight, are first obtained using a non-perturbative method. Then, a perturbation algorithm is used to analyze the target activation maps to obtain the first and second attention scores. The analysis focuses on the first and second feature activation maps that have a significant impact on the first and second feature activation maps, making it more targeted in obtaining the first and second attention scores, improving the calculation speed of the first and second attention scores, and ensuring the accuracy of the results.

[0155] The above describes some specific implementations of the image filtering method provided in this application. Based on this, this application also provides a corresponding apparatus. The apparatus provided in this application will be described below from the perspective of functional modularity.

[0156] Figure 2 The diagram shown is a structural schematic of an image filtering device. (See attached image file.) Figure 2 An image filtering device 200 includes:

[0157] The acquisition module 201 is used to acquire a library of images to be filtered, wherein the library of images to be filtered includes at least one image to be filtered;

[0158] The first processing module 202 is used to input the target image to be screened into the first target detection model, and obtain a first feature activation map set including multiple first feature activation maps through the first target detection model;

[0159] The second processing module 203 is used to input the target image to be screened into the second target detection model, and obtain a second feature activation map set including multiple second feature activation maps through the second target detection model. The accuracy of the second target detection model is higher than that of the first target detection model.

[0160] In one possible implementation, the first object detection model and the second object detection model are of the same type. The amount of data used to train the first object detection model is less than the amount of data used to train the second object detection model, and / or, the number of training iterations for the first object detection model is less than the number of training iterations for the second object detection model.

[0161] The first calculation module 204 is used to calculate the first weight of the first feature activation map in the first feature activation map set and the second weight of each second feature activation map in the second feature activation map set.

[0162] The second calculation module 205 is used to determine the first attention level of the first feature activation map set based on the feature values ​​and the first weight of the first feature activation map set.

[0163] The third calculation module 206 is used to determine the second attention level of the second feature activation map set based on the feature values ​​and second weights of the second feature activation map set.

[0164] The judgment module 207 is used to retain the target image to be filtered if the difference between the first attention and the second attention exceeds a preset complexity threshold; and to delete the target image to be filtered from the image library if the difference between the first attention and the second attention does not exceed the preset complexity threshold.

[0165] According to the aforementioned image filtering device, the same image to be filtered is processed using two target detection models with different capability levels. When analyzing complex images, if the two feature activation maps obtained by the two target detection models differ significantly, then the difference between the first and second attention levels is considered large, indicating high complexity of the image to be filtered. If the complexity exceeds a set complexity threshold, the image to be filtered is retained; otherwise, it is deleted from the image library. This achieves automatic filtering of images in the image library, improving filtering efficiency.

[0166] In one possible implementation, the first calculation module 204 is further configured to solve the first feature activation map set using Singular Value Decomposition (SVD) to obtain a right singular value matrix; and to use the values ​​of the first row of the right singular value matrix as the first weights of each first feature activation map in the first feature activation map set. Alternatively, the target image to be screened is input into a first target detection model to obtain the first initial data output by the first target detection model for identifying objects to be identified in the target image to be screened; the feature values ​​of the target first feature activation maps in the first feature activation map set are set to zero or a minimum value to obtain a perturbed first feature activation map set; the perturbed first feature activation map set is input into the data decision layer of the first target detection model to obtain the first perturbed data output by the data decision layer of the first target detection model; based on the first perturbed data and the first initial data, deviation data of the objects to be identified in the target image to be screened is obtained; and the first weight of the target first feature activation map is determined based on the deviation data.

[0167] The second calculation module 205 is further configured to multiply the feature value of each first feature activation map in the first feature activation map set by the first weight of the first feature activation map to obtain the feature weight value of the first feature activation map; and to sum the feature weight values ​​of the first feature activation maps included in the first feature activation map set to obtain the first attention of the first feature activation map set.

[0168] The first calculation module 204 is further configured to use Singular Value Decomposition (SVD) to solve the second feature activation map set to obtain a right singular value matrix; to use the values ​​of the first row of the right singular value matrix as the second weights of each second feature activation map in the second feature activation map set; or, to input the target image to be screened into a second target detection model to obtain the second initial data output by the second target detection model for identifying objects to be identified in the target image to be screened; to set the feature values ​​of the target second feature activation maps in the second feature activation map set to zero or a minimum value to obtain a perturbed second feature activation map set; to input the perturbed second feature activation map set into the data decision layer of the second target detection model to obtain the second perturbed data output by the data decision layer of the second target detection model; to obtain the deviation data of the objects to be identified in the target image to be screened based on the second perturbed data and the second initial data; and to determine the second weights of the target second feature activation maps based on the deviation data.

[0169] The third calculation module 206 is further configured to multiply the feature value of each second feature activation map in the second feature activation map set by the second weight of the second feature activation map to obtain the feature weight value of the second feature activation map; and to sum the feature weight values ​​of the second feature activation maps included in the second feature activation map set to obtain the second attention of the second feature activation map set.

[0170] In another possible implementation, the first calculation module 204 is further configured to obtain a first number of first feature activation maps included in the first feature activation map set;

[0171] If the first quantity exceeds a preset threshold, Singular Value Decomposition (SVD) is used to solve the first feature activation map set to obtain a right singular value matrix. The values ​​of the first row of the right singular value matrix are used as the weights of each first feature activation map in the first feature activation map set. The weights of all the first feature activation maps in the first feature activation map set are sorted, and a set number of first feature activation maps at the top of the sorting are selected to form a first target activation map set. The target image to be screened is input into a first target detection model to obtain the first initial data for identifying the object to be identified in the target image to be screened. The feature values ​​of the target first feature activation maps in the first target activation map set are set to zero or a minimum value to obtain a perturbed first target activation map set. The perturbed first target activation map set is input into the data decision layer of the first target detection model to obtain the first perturbed data output by the data decision layer of the first target detection model. According to the... The first perturbation data and the first initial data are used to obtain deviation data of the object to be identified in the target image to be screened; the first weight of the target first feature activation map is determined based on the deviation data; if the first quantity does not exceed a preset quantity threshold, the target image to be screened is input into a first target detection model to obtain the first initial data output by the first target detection model to identify the object to be identified in the target image to be screened; the feature values ​​of the target first feature activation map in the first feature activation map set are set to zero or a minimum value to obtain a perturbed first feature activation map set; the perturbed first feature activation map set is input into the data decision layer of the first target detection model to obtain the first perturbation data output by the data decision layer of the first target detection model; the deviation data of the object to be identified in the target image to be screened is obtained based on the first perturbation data and the first initial data; the first weight of the target first feature activation map is determined based on the deviation data.

[0172] The first calculation module 204 is further configured to: obtain a second number of second feature activation maps included in the second feature activation map set; if the second number exceeds a preset number threshold, use Singular Value Decomposition (SVD) to solve the second feature activation map set to obtain a right singular value matrix; use the values ​​of the first row of the right singular value matrix as the weights of each second feature activation map in the second feature activation map set; sort the weights of all the second feature activation maps in the second feature activation map set, and select a set number of second feature activation maps that are sorted first to form a second target activation map set; input the target image to be screened into the second target detection model to obtain the second initial data for identifying the object to be identified in the target image to be screened; set the feature values ​​of the target second feature activation maps in the second target activation map set to zero or a minimum value to obtain a perturbed second target activation map set; input the perturbed second target activation map set into the data decision layer of the second target detection model to obtain the second target detection... The model's data decision layer outputs second perturbation data; based on the second perturbation data and the second initial data, deviation data of the object to be identified in the target image to be screened is obtained; based on the deviation data, the second weight of the target second feature activation map is determined; if the second quantity does not exceed a preset quantity threshold, the target image to be screened is input into the second target detection model to obtain the second initial data output by the second target detection model to identify the object to be identified in the target image to be screened; the feature value of the target second feature activation map in the second feature activation map set is set to zero or a minimum value to obtain a perturbed second feature activation map set; the perturbed second feature activation map set is input into the data decision layer of the second target detection model to obtain the second perturbation data output by the data decision layer of the second target detection model; based on the second perturbation data and the second initial data, deviation data of the object to be identified in the target image to be screened is obtained; based on the deviation data, the second weight of the target second feature activation map is determined.

[0173] The second calculation module 205 is further configured to multiply the feature value of each first feature activation map to be analyzed in the first target activation map set by the first weight of the first feature activation map to be analyzed to obtain the feature weight value of the first feature activation map to be analyzed; sum the feature weight values ​​of the first feature activation maps to be analyzed included in the first target activation map set to obtain the first attention of the first target activation map set; and use the first attention of the first target activation map set as the first attention of the first feature activation map set.

[0174] The second calculation module 206 is further configured to multiply the feature value of each second feature activation map to be analyzed in the second target activation map set by the second weight of the second feature activation map to be analyzed to obtain the feature weight value of the second feature activation map to be analyzed; sum the feature weight values ​​of the second feature activation maps to be analyzed included in the second target activation map set to obtain the second attention of the second target activation map set; and use the second attention of the second target activation map set as the second attention of the second feature activation map set.

[0175] This application also provides corresponding devices and computer storage media for implementing the solutions provided in this application.

[0176] The device includes a memory and a processor. The memory is used to store instructions or code, and the processor is used to execute the instructions or code to cause the device to perform an image filtering method according to any embodiment of this application.

[0177] The computer storage medium stores code, and when the code is executed, the device running the code implements an image filtering method according to any embodiment of this application.

[0178] In the embodiments of this application, the terms "first" and "second" (if they exist) are used only as name identifiers and do not represent the order of first and second.

[0179] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus a general-purpose hardware platform. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as a read-only memory (ROM) / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a router) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0180] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0181] The above description is merely an exemplary implementation of this application and is not intended to limit the scope of protection of this application.

Claims

1. A picture screening method characterized by, The method comprises: obtaining a picture library to be screened, the picture library to be screened comprising at least one picture to be screened; inputting a target picture to be screened into a first target detection model, and obtaining a first feature activation map set comprising a plurality of first feature activation maps through the first target detection model; inputting the target picture to be screened into a second target detection model, and obtaining a second feature activation map set comprising a plurality of second feature activation maps through the second target detection model, the accuracy of the second target detection model being higher than that of the first target detection model; the first target detection model and the second target detection model are the same type of target detection model; calculating a first weight of each first feature activation map in the first feature activation map set and a second weight of each second feature activation map in the second feature activation map set; determining a first attention degree of the first feature activation map set according to the feature values of the first feature activation maps in the first feature activation map set and the first weights, comprising: multiplying the feature value of each first feature activation map in the first feature activation map set by the first weight of the first feature activation map to obtain a feature weight value of the first feature activation map; and summing the feature weight values of the first feature activation maps included in the first feature activation map set to obtain the first attention degree of the first feature activation map set; determining a second attention degree of the second feature activation map set according to the feature values of the second feature activation maps in the second feature activation map set and the second weights, comprising: multiplying the feature value of each second feature activation map in the second feature activation map set by the second weight of the second feature activation map to obtain a feature weight value of the second feature activation map; and summing the feature weight values of the second feature activation maps included in the second feature activation map set to obtain the second attention degree of the second feature activation map set; if the difference between the first attention degree and the second attention degree exceeds a preset complexity threshold, retaining the target picture to be screened; if the difference between the first attention degree and the second attention degree does not exceed the preset complexity threshold, deleting the target picture to be screened from the picture library to be screened.

2. The method of claim 1, wherein, The data amount for training the first target detection model is less than the data amount for training the second target detection model, and / or the training frequency of the first target detection model is less than the training frequency of the second target detection model.

3. The method of claim 1, wherein, The calculation of the first weight of each first feature activation map in the first feature activation map set comprises: solving the first feature activation map set by singular value decomposition (SVD) to obtain a right singular value matrix; taking the numerical values of the first row of the right singular value matrix as the first weights of each first feature activation map in the first feature activation map set; or inputting the target picture to be screened into the first target detection model to obtain first initial data for identifying the target to be identified in the target picture to be screened; setting the feature values of a target first feature activation map in the first feature activation map set to zero or a minimum value to obtain a perturbed first feature activation map set; ​ inputting the disturbed first feature activation map set into a data decision layer of the first target detection model, to obtain first disturbance data output by the data decision layer of the first target detection model; obtaining deviation data of the to-be-recognized object in the target to-be-screened picture according to the first disturbance data and the first initial data; determining the first weight of the target first feature activation map according to the deviation data.

4. The method of claim 1, wherein, The calculating the second weight of the second feature activation map in the second feature activation map set comprises: solving the second feature activation map set by singular value decomposition (SVD) to obtain a right singular value matrix; taking the numerical value of the first row of the right singular value matrix as the second weight of each second feature activation map in the second feature activation map set; or, inputting the target to-be-screened picture into the second target detection model to obtain second initial data output by the second target detection model for recognizing the to-be-recognized object in the target to-be-screened picture; setting the eigenvalue of the target second feature activation map in the second feature activation map set to zero or a minimum value, to obtain a disturbed second feature activation map set; inputting the disturbed second feature activation map set into a data decision layer of the second target detection model, to obtain second disturbance data output by the data decision layer of the second target detection model; obtaining deviation data of the to-be-recognized object in the target to-be-screened picture according to the second disturbance data and the second initial data; determining the second weight of the target second feature activation map according to the deviation data.

5. The method of claim 1, wherein, The calculating the first weight of the first feature activation map in the first feature activation map set comprises: obtaining a first number of first feature activation maps included in the first feature activation map set; if the first number exceeds a preset number threshold, solving the first feature activation map set by singular value decomposition (SVD) to obtain a right singular value matrix; taking the numerical value of the first row of the right singular value matrix as the weight of each first feature activation map in the first feature activation map set; sorting the weights of all first feature activation maps in the first feature activation map set, and selecting a set number of first feature activation maps in the front of the sorting to form a first target activation map set; inputting the target to-be-screened picture into the first target detection model to obtain first initial data output by the first target detection model for recognizing the to-be-recognized object in the target to-be-screened picture; setting the eigenvalue of the target first feature activation map in the first target activation map set to zero or a minimum value, to obtain a disturbed first target activation map set; inputting the disturbed first target activation map set into a data decision layer of the first target detection model, to obtain first disturbance data output by the data decision layer of the first target detection model; obtaining deviation data of the to-be-recognized object in the target to-be-screened picture according to the first disturbance data and the first initial data; determining the first weight of the target first feature activation map according to the deviation data. if the first number does not exceed the preset number threshold, inputting the target to-be-screened picture into the first target detection model to obtain first initial data output by the first target detection model for recognizing the to-be-recognized object in the target to-be-screened picture; set a feature value of a target first feature activation map in the first feature activation map set to zero or a minimum value, to obtain a disturbed first feature activation map set; input the disturbed first feature activation map set into a data decision layer of the first target detection model, to obtain first disturbance data output by the data decision layer of the first target detection model; obtain deviation data of the to-be-recognized object in the target picture to be screened according to the first disturbance data and the first initial data; determine a first weight of the target first feature activation map according to the deviation data.

6. The method of claim 5, wherein, The method further includes: multiply a feature value of each to-be-analyzed first feature activation map in the first target activation map set by the first weight of the to-be-analyzed first feature activation map to obtain a feature weight value of the to-be-analyzed first feature activation map; sum the feature weight values of the to-be-analyzed first feature activation maps included in the first target activation map set to obtain a first attention degree of the first target activation map set, and take the first attention degree of the first target activation map set as the first attention degree of the first feature activation map set.

7. The method of claim 1, wherein, The method further includes: obtain a second number of second feature activation maps included in the second feature activation map set; if the second number exceeds a preset number threshold, solve the second feature activation map set by singular value decomposition (SVD) to obtain a right singular value matrix; take a value of a first row of the right singular value matrix as a weight of each second feature activation map in the second feature activation map set; sort the weights of all the second feature activation maps in the second feature activation map set, and select a preset number of second feature activation maps in the front of the sorting to form a second target activation map set; input the target picture to be screened into the second target detection model to obtain second initial data output by the second target detection model for recognizing the to-be-recognized object in the target picture to be screened; set a feature value of a target second feature activation map in the second feature activation map set to zero or a minimum value, to obtain a disturbed second feature activation map set; input the disturbed second feature activation map set into a data decision layer of the second target detection model, to obtain second disturbance data output by the data decision layer of the second target detection model; obtain deviation data of the to-be-recognized object in the target picture to be screened according to the second disturbance data and the second initial data; determine a second weight of the target second feature activation map according to the deviation data. if the second number does not exceed the preset number threshold, input the target picture to be screened into the second target detection model to obtain second initial data output by the second target detection model for recognizing the to-be-recognized object in the target picture to be screened; set a feature value of a target second feature activation map in the second feature activation map set to zero or a minimum value, to obtain a disturbed second feature activation map set; input the disturbed second feature activation map set into a data decision layer of the second target detection model, to obtain second disturbance data output by the data decision layer of the second target detection model; According to the second disturbance data and the second initial data, deviation data of the to-be-recognized object in the target to-be-screened picture is obtained; According to the deviation data, a second weight of the target second feature activation map is determined.

8. The method of claim 7, wherein, determining the second attention degree of the second feature activation map set according to the feature values of the second feature activation maps in the second feature activation map set and the second weights comprises: multiplying the feature value of each to-be-analyzed second feature activation map in the second target activation map set by the second weight of the to-be-analyzed second feature activation map to obtain a feature weight value of the to-be-analyzed second feature activation map; summing the feature weight values of the to-be-analyzed second feature activation maps included in the second target activation map set to obtain the second attention degree of the second target activation map set, and taking the second attention degree of the second target activation map set as the second attention degree of the second feature activation map set.

9. An image screening apparatus characterized by comprising: comprises: an acquisition module configured to acquire a to-be-screened picture library, the to-be-screened picture library including at least one to-be-screened picture; a first processing module configured to input a target to-be-screened picture into a first target detection model, and obtain a first feature activation map set including a plurality of first feature activation maps through the first target detection model; a second processing module configured to input the target to-be-screened picture into a second target detection model, and obtain a second feature activation map set including a plurality of second feature activation maps through the second target detection model, the accuracy of the second target detection model being higher than that of the first target detection model; the first target detection model and the second target detection model are the same type of target detection model; a first calculation module configured to calculate first weights of the first feature activation maps in the first feature activation map set and second weights of the second feature activation maps in the second feature activation map set; a second calculation module configured to determine a first attention degree of the first feature activation map set according to feature values of the first feature activation maps in the first feature activation map set and the first weights, and comprising: multiplying the feature value of each first feature activation map in the first feature activation map set by the first weight of the first feature activation map to obtain a feature weight value of the first feature activation map; and summing the feature weight values of the first feature activation maps included in the first feature activation map set to obtain the first attention degree of the first feature activation map set; a third calculation module configured to determine a second attention degree of the second feature activation map set according to feature values of the second feature activation maps in the second feature activation map set and the second weights, and comprising: multiplying the feature value of each second feature activation map in the second feature activation map set by the second weight of the second feature activation map to obtain a feature weight value of the second feature activation map; and summing the feature weight values of the second feature activation maps included in the second feature activation map set to obtain the second attention degree of the second feature activation map set. A judgment module is configured to: if the difference between the first attention degree and the second attention degree exceeds a preset complexity threshold, retaining the target picture to be screened; and if the difference between the first attention degree and the second attention degree does not exceed the preset complexity threshold, deleting the target picture to be screened from the picture library to be screened.

Images