Method, system, electronic device and storage medium for determining false detection targets

Abstract

The present application discloses a method, system, electronic device and storage medium for determining falsely detected targets, the determination method comprising: determining the confidence of target detection features corresponding to each target among multiple targets detected from an image to be processed; clustering the first-category target detection features to obtain clustering results, wherein the first-category target detection features are target detection features whose confidence is greater than or equal to a preset confidence among target detection features corresponding to each target; based on the degree of deviation between each second-category target detection feature and the clustering result, determining the falsely detected targets among the targets corresponding to each second-category target detection feature; wherein the second-category target detection features include target detection features other than the first-category target detection features among the target detection features corresponding to each target. In the above manner, the present application can calculate the degree of deviation between the second-category target detection features and the clustering result to determine the falsely detected targets in the image to be processed.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method, system, electronic device and storage medium for determining a misdetected target. Background Art

[0002] Generally, in image processing, target detection refers to identifying the target or object of interest in an image, and then determining its location and other information. For example, in target detection, pedestrian detection often reduces the accuracy in order to ensure the recall rate, and identifies some objects in the background as target objects (such as people), which is called false detection, affecting the capture effect of human targets. Therefore, these false detection results need to be further removed to improve the accuracy of the final target detection.

[0003] However, among the existing technologies for determining false positives, most technologies require re-adjusting model training data or training classification models, which is a time-consuming process, has high data labeling costs, and requires high model generalization. Simply using a few layers of networks for false positive filtering is difficult to achieve a good accuracy rate. Summary of the invention

[0004] To solve the above-mentioned technical problems, the technical solution adopted in the first aspect of the present application is to provide a method for determining falsely detected targets, the method comprising: determining the confidence of a target detection feature corresponding to each target among multiple targets detected from an image to be processed; clustering the first-category target detection features to obtain clustering results, wherein the first-category target detection features are target detection features whose confidence is greater than or equal to a preset confidence among the target detection features corresponding to each target; based on the degree of deviation between each second-category target detection feature and the clustering result, determining the falsely detected targets among the targets corresponding to each second-category target detection feature; wherein the second-category target detection features include target detection features other than the first-category target detection features among the target detection features corresponding to each target.

[0005] In order to solve the above technical problems, the technical solution adopted in the second aspect of the present application is to provide a system for determining a misdetected target, comprising:

[0006] A determination module, used to determine the confidence of the target detection feature corresponding to each target among multiple targets detected from the image to be processed;

[0007] A clustering module, used for clustering the first type of target detection features to obtain clustering results, wherein the first type of target detection features are target detection features whose confidence is greater than or equal to a preset confidence among the target detection features corresponding to each target;

[0008] The determination module is further used to determine the misdetected targets among the targets corresponding to each second-category target detection feature based on the degree of deviation between each second-category target detection feature and the clustering result;

[0009] The second type of target detection features include target detection features corresponding to each target except the first type of target detection features.

[0010] In order to solve the above technical problems, the technical solution adopted in the third aspect of the present application is to provide an electronic device, which includes: a processor and a memory, the memory stores a computer program, and the processor is used to execute the computer program to implement the determination method as in the first aspect of the present application.

[0011] In order to solve the above technical problems, the technical solution adopted in the fourth aspect of the present application is to provide a computer-readable storage medium, which stores a computer program, and the computer program can implement the determination method of the first aspect of the present application when executed by a processor.

[0012] The beneficial effects of the present application are as follows: the present application first screens the target detection features corresponding to the preset confidence level, thereby narrowing the scope of false detection retrieval required; and clusters the first-category target detection features corresponding to the preset confidence level, thereby determining the comparison standard for the second-category target detection features; based on the degree of deviation between each second-category target detection feature and the clustering result, the falsely detected targets in the image can be accurately identified, thereby determining the detected targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0014] Figure 1 This is a flow chart of a first embodiment of a method for determining a misdetected target of the present application;

[0015] Figure 2 This is a schematic diagram of the process of extracting detection frame features using a fast neural network in this application;

[0016] Figure 3 yes Figure 1 A schematic diagram of a specific implementation process of step S13;

[0017] Figure 4 yes Figure 3 A schematic diagram of a specific implementation process of step S21;

[0018] Figure 5 It is a flowchart of a specific embodiment of the present application for obtaining cluster centers in the first type of target detection features;

[0019] Figure 6 This is a flowchart of a specific embodiment of iteratively updating the cluster center of the present application;

[0020] Figure 7 It is a flowchart of another specific embodiment of iteratively updating the cluster center of the present application;

[0021] Figure 8 It is a schematic block diagram of the structure of an embodiment of a system for determining a misdetected target of the present application;

[0022] Fig. 9 is a schematic structural block diagram of an embodiment of an electronic device of the present application;

[0023] Fig.10 It is a circuit schematic block diagram of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0024] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0025] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0026] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0027] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0028] As used in this specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0029] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0030] In order to illustrate the technical solution of the present application, the following is a specific embodiment to illustrate that the present application provides a method for determining a misdetected target. Figure 1 , Figure 1 1 is a flow chart of a first embodiment of a method for determining a misdetected target of the present application, and the method specifically comprises the following steps:

[0031] S11: determining the confidence of the target detection feature corresponding to each target among multiple targets detected from the image to be processed;

[0032] Generally speaking, for a certain scene, a video image including multiple targets, such as people, motor vehicles and non-motor vehicles, can be captured, and multiple targets can be detected from the image to be processed. Usually, a video image is formed by multiple frames of pictures, and the representation of the target in each frame is different. Confidence confirmation is required to confirm whether it is a real target and distinguish it from the environment. The image to be processed can be multiple targets detected from multiple images extracted from the video (mainly this scene), or multiple targets detected in one image, which is not limited here.

[0033] Usually, target detection features refer to the image features of the image area covered by the detected target box. To obtain a video image of a certain scene, a picture stream can be extracted from the video at a fixed frame rate, and each picture can be input into the target detection model for image feature detection to obtain the output target detection features. The target detection features include target type, target box size, and confidence, so that the confidence of the target detection features corresponding to each target can be determined.

[0034] Specifically, for example, the target type detected in a certain image is a motor vehicle; the target box is the circumscribed rectangular box of the motor vehicle; the confidence level is the confidence level of the target detection model in identifying the detection box as a target, and the value range is [0,1], which means whether the detected target is a real target.

[0035] S12: clustering the first type of target detection features to obtain a clustering result;

[0036] Usually, the determination method is provided with a preset confidence level, which is used as a criterion for judging whether it is a real target. The confidence level of each target detection feature is compared with the preset confidence level, so as to determine that the confidence level of the first type of target detection feature is greater than or equal to the preset confidence level, and determine that the confidence level of the second type of target detection feature is less than the preset confidence level. That is, the second type of target detection feature includes target detection features corresponding to each target except the first type of target detection features.

[0037] Specifically, assuming that the preset confidence is 0.8, if there are 100 confidences of target detection features, of which 60 have confidences greater than or equal to 0.8, it means that the confidences of 60 first-category target detection features are high confidences. If the confidences of the remaining 40 target detection features are less than 0.8, it means that the confidences of these 40 second-category target detection features are low confidences.

[0038] In order to quickly identify falsely detected targets, the first type of target detection features can be clustered. Specifically, the target detection features corresponding to the confidence level greater than or equal to the preset level are clustered. The purpose is to find the clustering results, which will serve as a reference standard for judging the second type of target detection features with a confidence level less than the preset level.

[0039] Specifically, as in the example described in step S12, the target detection features extracted from 60 target frames with higher confidence levels are clustered and initialized, and used as a reference standard for the target detection features extracted from 40 target frames with lower confidence levels.

[0040] S13: determining misdetected targets among targets corresponding to each second-category target detection feature based on the degree of deviation between each second-category target detection feature and the clustering result;

[0041] Due to the limited shooting conditions and technology, in some scenes, the target and the scene are integrated into one, and the detected target is prone to false detection, which affects the main capture effect of the target. In order to find out the falsely detected target, specifically, based on the degree of deviation between each second-class target detection feature and the clustering result, the falsely detected target among the targets corresponding to each second-class target detection feature can be determined, wherein the greater the degree of deviation, the target corresponding to the second-class target detection feature is a falsely detected target, and the smaller the degree of deviation, the target corresponding to the second-class target detection feature is a real target.

[0042] Therefore, the present application first screens the target detection features corresponding to the preset confidence level that is greater than or equal to the preset confidence level, thereby narrowing the scope of false detection retrieval required; and clusters the first-category target detection features corresponding to the preset confidence level that is greater than or equal to the preset confidence level, and determines the comparison standard for the second-category target detection features; based on the degree of deviation between each second-category target detection feature and the clustering result, it is possible to accurately identify the falsely detected targets in the image, thereby determining the falsely detected targets.

[0043] Among them, for target detection in the acquired image, a target detection model is needed, and feature box detection and extraction of the target are performed through the target detection model. Among them, the target detection model can use a fast neural network (Faster Region Convolutional Neural Network, Faster RCNN) to directly extract the features of the detection box from the feature output layer of Faster RCNN, thereby obtaining the target detection feature.

[0044] Specifically, see Figure 2 , Figure 2 This is a processing diagram of the fast neural network used in this application to extract detection frame features. First, the image is input into the convolutional layers (Conv layers) in Faster RCNN to extract the target features of the entire image; then the target features are input into the Region Proposal Networks (RPN) to generate candidate frames; and then the candidate frames are corrected using bounding box regression to obtain an accurate detection frame.

[0045] These detection boxes are then mapped to the output layer of the Convolutional Neural Network (CNN) to obtain the region of interest (ROI) and the features of the mapping area. These ROI features are used as the input of the ROI Pooling layer. After the ROI pooling operation, the features of the detection box of a fixed size are obtained, so that the target can be subsequently removed from false detection.

[0046] Furthermore, based on the degree of deviation between each second-class target detection feature and the clustering result, the number of falsely detected targets among the targets corresponding to each second-class target detection feature is determined. Figure 3 , Figure 3 yes Figure 1 A specific implementation flow diagram of step S13 in the embodiment of the present invention specifically includes the following steps:

[0047] S21: Based on the clustering result, respectively determine the detection error value of each second-category target detection feature;

[0048] The detection error value is determined based on the degree of deviation between the corresponding second-category target detection features and the clustering results.

[0049] Specifically, the clustering result and each second-category target detection feature may be input into a detection error function, and the degree of deviation between each corresponding second-category target detection feature and the clustering result may be calculated, thereby calculating and determining the detection error value of each second-category target detection feature.

[0050] S22: Determine the targets corresponding to the second-category target detection features whose detection error values ​​are greater than or equal to the detection error value threshold as falsely detected targets.

[0051] The determination method is also provided with a detection error value threshold value, which is used to check whether the deviation between the corresponding second-category target detection features and the clustering result exceeds the allowed range. If it exceeds the allowed range, it means that the target corresponding to the second-category target detection feature is a false detection target. If it does not exceed the allowed range, it means that the target corresponding to the second-category target detection feature is a true target. Therefore, when the detection error value is greater than or equal to the detection error value threshold value, the target corresponding to each second-category target detection feature whose detection error value is approximately or equal to the detection error value threshold value can be determined as the false detection target.

[0052] Furthermore, based on the clustering results, the detection error values ​​of each second-class target detection feature are determined respectively. Figure 4 , Figure 4 yes Figure 3 A specific implementation flow diagram of step S21 in the embodiment of the present invention specifically includes the following steps:

[0053] S31: Determine each second-category target detection feature that has a confidence level less than a preset confidence level;

[0054] The confidences of the multiple target detection features are compared with the preset confidences respectively, and a judgment method can be adopted. When it is determined that the target detection feature is less than the preset confidence, each second type target detection feature less than the preset confidence is determined.

[0055] For example: select 100 target detection frames from target detection done by others, and compare the confidence of the 100 target detection frames with the preset confidence. Among them, the confidence of 40 target detection frames is less than the preset confidence. Then it can be determined that the target detection features corresponding to the confidence of the 40 target detection frames are the second-category target detection features that are less than the preset confidence.

[0056] S32: Input the clustering results and each second-category target detection feature into a cosine function, and calculate a plurality of cosine distances to serve as detection error values.

[0057] Specifically, the detection error function may be a cosine function, and the detection error value may be a cosine distance. Therefore, the clustering result and the second-category target detection feature are both input into the cosine function to obtain the cosine distance.

[0058] Since there are multiple second-category target detection features, each time the clustering result and a second-category target detection feature are input into the cosine function to calculate a cosine distance. By performing the same input operation multiple times, multiple cosine distances can be obtained as detection error values.

[0059] Furthermore, if the clustering result is a clustering center, the first-category target detection feature is clustered to obtain a clustering result. Therefore, specifically, based on multiple clustering centers, combined with the loss function and the cosine function, the first-category target detection feature can be clustered to obtain the clustering center.

[0060] For further information, see Figure 5 , Figure 5 yes Figure 5 This is a schematic diagram of a specific embodiment of the present application for obtaining cluster centers in the first type of target detection features. Based on multiple cluster centers, the first type of target detection features are clustered in combination with the loss function and the cosine function to obtain the cluster centers, which specifically includes the following steps:

[0061] S41: Randomly select a target detection feature from the first type of target detection features as a cluster center;

[0062] Specifically, in the first category of target detection features, a target detection feature can be randomly selected as a cluster center to initialize a cluster center C0.

[0063] S42: Based on the loss function, calculate the loss cost between the cluster center and the remaining target detection features;

[0064] Specifically, after determining the cluster center C0, we then calculate the cluster center C j X i The cosine distance dist(Cj ,X i ), the loss function is d j It is expressed as formula (1):

[0065]

[0066] Wherein, j is a positive integer greater than or equal to 0, and its maximum number is the number of the first type of target detection features; i is a positive integer greater than or equal to 1, and its maximum number is the number of the second type of target detection features; the value of N is the number of the second type of target detection features, X i Indicates target detection features that are less than a preset confidence level.

[0067] Thus, the loss cost value of the cluster center can be calculated, which provides a comparison value for selecting a suitable cluster center in the first category target detection feature in the future.

[0068] S43: Based on the iterative function, combined with the loss cost value and the cluster center, iterative calculation is performed to obtain the updated cluster center;

[0069] Specifically, combined with the loss cost d j and the previous cluster center C j , the previous cluster center C j Input the iterative function, which is shown in formula (2):

[0070]

[0071] Among them, η is the learning rate. By increasing j continuously, the cluster center can change. Through the iterative function, the cluster center C j Shift to the real cluster center and get the new cluster center C j+1 .

[0072] S44: Determine whether the loss cost is the minimum;

[0073] Specifically, the cluster center of the first type of target detection features is updated iteratively, and multiple loss cost values ​​of the cluster center and the remaining target features are calculated. Then, the multiple loss cost values ​​are compared to determine whether the loss cost value is the smallest.

[0074] If not, it means that the loss cost value is not the minimum and iterative calculation needs to continue, then return to step S42, that is, based on the loss function, calculate the loss cost value between the cluster center and the remaining target detection features; if yes, it means that the minimum loss cost value has been found and no further iterative update is required, then enter step S45, that is, stop iteration and determine that the target detection feature corresponding to the minimum loss cost value is the cluster center.

[0075] Therefore, by performing multiple clustering and selecting the one with the smallest final loss cost, the true cluster center of the first type of target detection features can be obtained.

[0076] Furthermore, for the multiple target detection features obtained, the part greater than or equal to the preset confidence is classified into the first category of target detection features for finding the cluster center, and the part less than the preset confidence is classified into the second category of target detection features for planning the calculation range for the subsequent calculation of the deviation degree between each corresponding second category target detection feature and the cluster center.

[0077] Specifically, the cluster centers of the first type of target detection features and the second type of target detection features are input into the cosine distance function, and the cosine distance d between the cluster centers of the second type of target detection features and the first type of target detection features can be calculated respectively. i , the cosine distance function is shown in formula (3):

[0078] d i =1-cos <Y i ,C> (3)

[0079] Among them, Y i represents the second type of target detection features, and C represents the true cluster center in the model sample.

[0080] Furthermore, after determining the falsely detected targets among the targets corresponding to each second-category target detection feature, the method further includes: outputting detection information of targets other than the falsely detected targets among the multiple targets.

[0081] For further information, see Figure 6 , Figure 6 : is a schematic diagram of a specific embodiment of iteratively updating the cluster center of the present application, and the method also includes:

[0082] S51: Determine whether the cosine distance is less than the detection error value threshold;

[0083] Specifically, let the detection error value threshold be γ. If γ is set according to specific circumstances and actual experience, it is used to compare with the cosine distance to determine whether the cosine distance is less than the detection error value threshold.

[0084] If it is less than the detection error threshold, it means d i <γ, indicating that the target corresponding to the target detection feature is a real target, then the process goes to step S52, that is, adding the target corresponding to the value less than the detection error threshold to the first type of target detection feature, and Figure 5 The embodiment in is calculated to iteratively update the cluster center, thereby obtaining a new cluster center C1.

[0085] If it is greater than or equal to the detection error value threshold, it means di ≥γ, the object in the detection frame corresponding to the target detection feature is a false detection target, and the process goes to step S53, that is, determining the false detection target.

[0086] For further information, see Figure 7 , Figure 7 1 is another specific embodiment of the present invention, which is a flow chart of iteratively updating the cluster center. In this embodiment, it is a way of incrementally iterating the cluster center with confidence target. The method specifically includes the following steps:

[0087] S61: Obtain a detection frame corresponding to the target as a new target detection feature;

[0088] Specifically, for the target in the subsequent video image, the image stream can be extracted from the video at a fixed frame rate to obtain its detection result, that is, to obtain the detection frame corresponding to the target. Specifically, Figure 2 The steps performed in are used to obtain it, which will not be repeated here.

[0089] S62: Determine whether the confidence level corresponding to the detection frame is greater than or equal to a preset confidence level;

[0090] Specifically, it is possible to determine whether the confidences corresponding to the detection boxes are greater than the preset confidences by one-to-one comparison, or to determine whether the confidences corresponding to the detection boxes are greater than the preset confidences by asynchronously performing comparisons in multiple groups, and then retain the detection boxes with confidences greater than the preset confidences.

[0091] Specifically, for multiple groups of asynchronous comparisons, if there are 100 confidence levels, assuming the preset confidence level is 0.8, the 100 confidence levels can be divided into 5 groups, each including 20 confidence levels. Then the confidence levels in each group are compared with the preset confidence levels, and the confidence levels in each group that are greater than or equal to 0.8 can be obtained. In this way, multiple groups of asynchronous comparisons are formed, and the groups do not interfere with each other. Confidence levels greater than or equal to 0.8 can be obtained among the 100 confidence levels. Therefore, the confidence levels greater than or equal to 0.8 can be found more quickly, thereby improving comparison efficiency.

[0092] S63: Add the new target detection features to the first type of target detection features to iteratively update the cluster center.

[0093] Specifically, for detection boxes with higher confidence after detection, the target detection features can be directly extracted from the feature output layer of the detection network Faster RCNN and added to the model sample. Then the initial cluster center is C1, and the cluster center C is calculated. j The object detection feature S of each high-confidence object box i The cosine distance dist(C j ,S i ),

[0094]

[0095] Wherein, j is a positive integer greater than or equal to 0, and its maximum number is the number of target detection features greater than or equal to the preset confidence; i is a positive integer greater than or equal to 1, and its maximum number is the number of target detection features less than the preset confidence; the value of m is the number of updates of target detection features greater than or equal to the preset confidence after removing the cluster center, S i Represents the remaining confidence of the object detection features after removing the cluster center.

[0096] Furthermore, the cluster center can be obtained according to the iterative function, and its iterative function is shown in formula (5):

[0097]

[0098] Among them, δ is the learning rate; by continuously increasing j, the cluster center can change, and through the iterative function, the original cluster center can be shifted to the new real cluster center to obtain a new cluster center.

[0099] For newly detected high-confidence target detection features, the model samples are updated in an online, incremental iterative method according to the above steps. For newly detected low-confidence target detection features, they are sequentially input into the model samples for intermediate judgment.

[0100] Therefore, this application further judges the false detections generated after target detection and establishes a model for quickly removing false detections online. This model is a false detection removal model that can adapt to the scene and realize incremental updates. This process does not require long-term training, and the model is simple and fast. In addition, the online method can collect and transmit data in real time, which is convenient for data storage and analysis. Fast, online, and incremental update iterations allow model samples to update the cluster center in real time according to the newly added data, and the model samples gradually become stable and accurate.

[0101] In this way, it is possible to establish a model sample online to quickly remove false positives. As the number of target samples increases, the model will iteratively update the cluster center, which is an incremental update false positive removal method. In addition, it does not require long-term training, saving the manpower and time costs required for labeling and training data, and improving the speed and efficiency of false positive removal. It has the advantages of saving time, being simple and convenient, and being highly practical. Moreover, considering the multi-type and multi-scenario detection in target detection, including but not limited to pedestrian detection, motor vehicle detection, etc., it is a false positive removal method that can adapt to needs and scenarios.

[0102] In order to illustrate the technical solution of the present application, the present application also provides a system for determining a misdetected target, which can be installed on a computer, a server, or a mobile terminal, and has an online detection function, which is not limited here. Figure 8 , Figure 8 It is a schematic structural block diagram of an embodiment of a system for determining a falsely detected target of the present application. The system for determining a falsely detected target includes a determination module 71 and a clustering module 72 .

[0103] A determination module 71, used to determine the confidence level of the target detection feature corresponding to each target among multiple targets detected from the image to be processed;

[0104] The clustering module 72 is used to cluster the first type of target detection features to obtain clustering results; the first type of target detection features are target detection features whose confidence is greater than or equal to a preset confidence among the target detection features corresponding to each target;

[0105] The determination module 71 is also used to determine the falsely detected targets among the targets corresponding to each second-category target detection feature based on the degree of deviation between each second-category target detection feature and the clustering result; the second-category target detection features include the target detection features corresponding to each target except the first-category target detection features.

[0106] Therefore, the present application first screens and determines the first-category target detection features corresponding to the preset confidence level by the determination module 71, thereby narrowing the scope of false detection retrieval required; and initially clusters the second-category target detection features corresponding to the preset confidence level by the clustering module 72, thereby determining the target detection feature comparison standard; based on the degree of deviation between each second-category target detection feature and the clustering result, the falsely detected targets in the image can be accurately identified, thereby determining the falsely detected targets.

[0107] In order to illustrate the technical solution of the present application, the present application also provides an electronic device, which can be a computer or a mobile phone, etc., without specific limitation. Fig. 9 , Fig. 9 It is a schematic block diagram of the structure of an electronic device embodiment of the present application. The electronic device 8 includes: a processor 81 and a memory 82. The memory 82 stores a computer program 821. The processor 81 is used to execute the computer program 821 to implement the determination method of the first aspect of the embodiment of the present application, which will not be repeated here.

[0108] In addition, this application also provides a computer-readable storage medium, see Fig.10 , Fig.10It is a circuit schematic block diagram of an embodiment of a computer-readable storage medium of the present application. The computer-readable storage medium 9 stores a computer program 91. When the computer program 91 is executed by a processor, it can implement the determination method of the first aspect of the embodiment of the present application, which will not be repeated here.

[0109] If it is implemented in the form of a software functional unit and sold or used as an independent product, it can also be stored in a device with a storage function. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage device, including a number of instructions (program data) to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of each implementation method of the present invention. The aforementioned storage device includes: various media such as USB flash drive, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and electronic devices such as computers, mobile phones, laptops, tablet computers, cameras, etc. with the above storage media.

[0110] The description of the execution process of the program data in the device with storage function can be referred to the description in the above-mentioned embodiment of the method for determining the false detection target of the present application, which will not be repeated here.

[0111] The above descriptions are merely embodiments of the present application and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for determining a misdetected target, characterized in that: The determination method comprises: Determining the confidence level of the target detection feature corresponding to each target among multiple targets detected from the image to be processed; Clustering the first type of target detection features to obtain clustering results, wherein the first type of target detection features are target detection features whose confidence is greater than or equal to a preset confidence among the target detection features corresponding to the respective targets; Determining misdetected targets among targets corresponding to each second-category target detection feature based on the degree of deviation between each second-category target detection feature and the clustering result; The second type of target detection features includes target detection features corresponding to the respective targets except for the first type of target detection features.

2. The determination method according to claim 1, characterized in that: The determining, based on the degree of deviation between each second-category target detection feature and the clustering result, the misdetected target among the targets corresponding to each second-category target detection feature comprises: Based on the clustering result, respectively determine the detection error value of each second-category target detection feature, wherein the detection error value is determined based on the degree of deviation between the corresponding each second-category target detection feature and the clustering result; The targets corresponding to the second-category target detection features whose detection error values ​​are greater than or equal to the detection error value threshold are determined as the misdetected targets.

3. The determination method according to claim 2, characterized in that: The step of respectively determining the detection error values ​​of the respective second-category target detection features based on the clustering result includes: Determining each of the second-category target detection features having a confidence level less than the preset confidence level; The clustering result and each of the second-category target detection features are input into a cosine function, and a plurality of cosine distances are calculated to be used as the detection error value.

4. The determination method according to claim 3, characterized in that: The clustering result is the cluster center; The step of clustering the first type of target detection features to obtain clustering results includes: Based on multiple cluster centers, combined with the loss function and the cosine function, the first category target detection features are clustered to obtain the cluster centers.

5. The determination method according to claim 4, characterized in that: The clustering of the first type of target detection features based on multiple cluster centers, combined with a loss function and a cosine function, to obtain the cluster centers includes: Randomly selecting a target detection feature from the first type of target detection features as the cluster center; Based on the loss function, calculating the loss cost between the cluster center and the remaining target detection features; Based on the iterative function, combining the loss cost value and the cluster center, iteratively calculating to obtain an updated cluster center, and returning to the step of calculating the loss cost value between the cluster center and the remaining target detection features based on the loss function, until the minimum loss cost value is obtained and then stopping the iteration; Determine the target detection feature corresponding to the minimum loss cost value as the cluster center.

6. The determination method according to claim 1, characterized in that: After determining the misdetected targets among the targets corresponding to the second-category target detection features, the method further includes: Detection information of targets other than the misdetected target among the plurality of targets is output.

7. The determination method according to claim 4, characterized in that: The determination method further comprises: Determine whether the cosine distance is less than a detection error value threshold; If it is less than the detection error value threshold, the target detection features corresponding to the values ​​less than the detection error value threshold are added to the first category of target detection features and the cluster center is iteratively updated.

8. The method according to claim 4, characterized in that The method further comprises: Get the detection box corresponding to the target as a new target detection feature; If it is determined that the confidence corresponding to the detection box is greater than or equal to the preset confidence, the new target detection feature is added to the first type of target detection feature to iteratively update the cluster center.

9. A system for determining a misdetected target, characterized in that: include: A determination module, used to determine the confidence of the target detection feature corresponding to each target among multiple targets detected from the image to be processed; A clustering module, used for clustering the first type of target detection features to obtain a clustering result; the first type of target detection features are target detection features whose confidence is greater than or equal to a preset confidence among the target detection features corresponding to the respective targets; The determination module is also used to determine the falsely detected targets among the targets corresponding to each second-category target detection feature based on the degree of deviation between each second-category target detection feature and the clustering result; the second-category target detection features include the target detection features corresponding to each target except the first-category target detection features.

10. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores a computer program, and the processor is used to execute the computer program to implement the determination method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the determination method according to any one of claims 1 to 8 can be implemented.

Images