Feature extraction method, electronic equipment and storage medium

Abstract

The invention discloses a feature extraction method, electronic equipment and a storage medium. The feature extraction method comprises the steps of obtaining a to-be-extracted image and a target scene type corresponding to the to-be-extracted image, wherein the target scene type is used for representing a scene type of an image acquisition scene where the to-be-extracted image is located; selecting a candidate detection operator matched with the target scene type from a plurality of candidate detection operators in the target detection model as a target detection operator, wherein different candidate detection operators can perform target detection on the same type of target objects under different scene types; performing target detection on the to-be-extracted image by using the target detection operator to obtain a target detection result corresponding to the to-be-extracted image, the target detection result comprising a plurality of candidate detection frames; and performing feature extraction on the plurality of candidate detection frames to obtain a feature extraction result of the to-be-extracted image. According to the scheme, the accuracy of a feature extraction result can be improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a feature extraction method, an electronic device, and a storage medium. Background Art

[0002] In recent years, feature extraction can assist in understanding and interpreting the content in images or videos, and feature extraction has been widely used in computer vision processing. Feature extraction can support various applications, such as target retrieval. After the image acquisition device acquires the image data, it usually performs target detection on the image data, and then performs feature extraction. Specifically, the specific position of the target object in the image is first found through the target detection algorithm, and then the contour information and key part information of the target object are extracted to generate the feature extraction result. In practical applications, because different images may have different image acquisition scenes, the target detection operator of a single image acquisition scene cannot take into account the image acquisition scenes of all images. Images of different acquisition scenes only use the target detection operator suitable for one image acquisition scene, which may result in inaccurate target detection results of the target detection operator. Furthermore, inaccurate target detection results often directly affect the results of subsequent feature extraction, and feature extraction failure may occur.

[0003] In view of the existing technical defects, how to provide a solution with more accurate feature extraction results is a technical problem that needs to be solved urgently by technical personnel in this field. Summary of the invention

[0004] The present application at least provides a feature extraction method, an electronic device and a storage medium.

[0005] The present application provides a feature extraction method, including: obtaining an image to be extracted and a target scene type corresponding to the image to be extracted, the target scene type being used to indicate the scene type of an image acquisition scene in which the image to be extracted is located; selecting a candidate detection operator that matches the target scene type from a plurality of candidate detection operators in a target detection model as a target detection operator, wherein different candidate detection operators can perform target detection on target objects of the same type under different scene types; performing target detection on the image to be extracted using the target detection operator to obtain a target detection result corresponding to the image to be extracted, wherein the target detection result includes a plurality of candidate detection frames; performing feature extraction on the plurality of candidate detection frames to obtain a feature extraction result of the image to be extracted.

[0006] The present application provides a feature extraction device, comprising: an acquisition module, a selection module, a target detection module and a feature extraction module; the acquisition module is used to acquire an image to be extracted and a target scene type corresponding to the image to be extracted, the target scene type is used to indicate the scene type of an image acquisition scene in which the image to be extracted is located; the selection module is used to select a candidate detection operator that matches the target scene type from a plurality of candidate detection operators in a target detection model as a target detection operator, and different candidate detection operators can perform target detection on target objects of the same type under different scene types; the target detection module is used to perform target detection on the image to be extracted using the target detection operator, and obtain a target detection result corresponding to the image to be extracted, and the target detection result includes a plurality of candidate detection frames; the feature extraction module is used to perform feature extraction on the plurality of candidate detection frames, and obtain a feature extraction result of the image to be extracted.

[0007] The present application provides an electronic device, including a memory and a processor, wherein the processor is used to execute program instructions stored in the memory to implement the above-mentioned feature extraction method.

[0008] The present application provides a computer-readable storage medium on which program instructions are stored. When the program instructions are executed by a processor, the above-mentioned feature extraction method is implemented.

[0009] The above scheme obtains the image to be extracted and the target scene type corresponding to the image to be extracted, and uses the target detection operator matching the target scene type to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted. Compared with the prior art that uses the same target detection operator to perform target detection on the same type of target objects under different scene types, the present application can improve the accuracy of the target detection result, thereby performing feature extraction on several candidate detection boxes in the target detection result, and obtaining the feature extraction result of the image to be extracted, which can improve the accuracy of the feature extraction result of the image to be extracted.

[0010] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0012] Figure 1 This is a schematic diagram of the process of an embodiment of the feature extraction method of the present application. Figure 1 ;

[0013] Figure 2 This is a schematic diagram of the process of an embodiment of the feature extraction method of the present application. Figure 2 ;

[0014] Figure 3 This is a schematic diagram of the process of an embodiment of the feature extraction method of the present application. Figure 3 ;

[0015] Figure 4 It is a structural schematic diagram of an embodiment of a feature extraction device of the present application;

[0016] Figure 5 It is a structural schematic diagram of an embodiment of the electronic device of the present application;

[0017] Figure 6 It is a structural diagram of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0018] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0019] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0020] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the objects associated before and after are in an "or" relationship. In addition, "many" in this article means two or more than two. In addition, the term "at least one" in this article means any combination of at least two of any one or more of a plurality of, for example, including at least one of A, B, and C, can mean including any one or more elements selected from the set consisting of A, B, and C.

[0021] The present application provides some feature extraction methods and feature extraction devices. The application scenarios of the feature extraction method include but are not limited to feature extraction of images collected by an image acquisition device. The executor of the feature extraction method may be a feature extraction device. For example, the feature extraction device may be arranged in a terminal device or a server or other processing device, wherein the terminal device may be a device for feature extraction, a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, etc. In some possible implementations, the feature extraction method may be implemented by a processor calling computer-readable instructions stored in a memory.

[0022] See also Figure 1 , Figure 1This is a schematic diagram of the process of an embodiment of the feature extraction method of the present application. Figure 1 .

[0023] Specifically, the feature extraction method may include the following steps:

[0024] Step S11: Acquire the image to be extracted and the target scene type corresponding to the image to be extracted.

[0025] The target scene type is used to indicate the scene type of the image acquisition scene where the image to be extracted is located.

[0026] The image to be extracted may be an image captured of the same type of target object under different scene types. The target object may be any object that needs to be feature extracted, including but not limited to animals and vehicles. Exemplarily, target object A and target object B both belong to the target class in the animal category, and the target class may be a dog or a cat. Then, the same type of target objects may be understood as target object A and target object B may be different dogs or different cats. The image to be extracted may be an image captured of the target area of ​​the same type of target object under different scene types. The target area may be the area corresponding to the face of the target object, that is, the facial area, or the area corresponding to the body of the target object, that is, the body area. It is understandable that the present application does not limit the type to which the target object belongs. The present application takes the type of the target object as a cat as an example, and takes the target area as the area corresponding to the face of the target object as an example.

[0027] The scene type of the image acquisition scene where the image to be extracted is located is the target scene type. The target scene type can be one of several candidate scene types, or can be other scene types other than several candidate scene types. Each candidate scene type can be a preset scene type.

[0028] Exemplarily, the several candidate scene types include a first candidate scene type, a second candidate scene type, and a third candidate scene type. The first candidate scene type may be scene type A, the second candidate scene type may be scene type B, and the first candidate scene type may be scene type C. Specifically, scene type A may represent a facial image acquisition scene of a cat captured in a car window. Scene type B may represent a facial image acquisition scene of a cat in an ID photo acquisition scene. Scene type C may represent a facial image acquisition scene of a cat in a video frame. In some application scenarios, the target scene type corresponding to the image to be extracted may be the first candidate scene, that is, the target scene type corresponding to the image to be extracted obtained in step S11 may be scene type A. In other application scenarios, the target scene type corresponding to the image to be extracted may be other scene types. Other scene types are not any one of the several candidate scene types, that is, the target scene type corresponding to the image to be extracted obtained in step S11 may also be scene type D, and scene type D is other scene types other than the several candidate scene types.

[0029] Step S12: Selecting a candidate detection operator that matches the target scene type from a number of candidate detection operators in the target detection model as the target detection operator.

[0030] Different candidate detection operators can perform target detection on the same type of target objects in different scene types.

[0031] The target detection model includes several candidate detection operators. Among them, the target detection model includes several candidate detection operators set in parallel. Each candidate detection operator can perform target detection processing on the images to be extracted of the same type of target objects belonging to different scene types. The target detection model is obtained through pre-training. Each candidate detection operator in the trained target detection model can perform target detection on the images to be extracted collected in a specific image acquisition scene, and obtain better target detection effect. In some application scenarios, each candidate detection operator in the target detection model can collect a scene type that matches each candidate detection operator for a specific image acquisition scene.

[0032] The scene types that match each candidate detection operator are candidate scene types, and the total number of each candidate scene type is the same as the total number of each candidate detection operator. Each candidate scene type can be a first candidate detection operator matching a first candidate scene type, a second candidate detection operator matching a second candidate scene type, and a third candidate detection operator matching a third candidate scene type. Specifically, the first candidate detection operator can be bound to scene type A, the second candidate detection operator can be bound to scene type B, and the third candidate detection operator can be bound to scene type C. Each candidate detection operator can perform target detection on any image to be extracted that is collected under the scene type. In other application scenarios, each candidate detection operator can perform target detection on images to be extracted that are collected in other image collection scenes other than a specific image collection scene. It can be understood that different candidate detection operators can all be operators for performing target detection on target objects of the same type, and the specific image collection scenes matched by each candidate detection operator are different scene types. Specifically, different candidate detection operators can all be operators for performing target detection on the same preset parts of target objects of the same type. Exemplarily, the target detection model includes several candidate detection operators, namely, a first candidate detection operator, a second candidate detection operator, and a third candidate detection operator. Among them, the first candidate detection operator may be detection operator a. The second candidate detection operator may be detection operator b. The third candidate detection operator may be detection operator c. It is understandable that there is a matching relationship between detection operator a and scene type A, a matching relationship between detection operator b and scene type B, and a matching relationship between detection operator c and scene type C. Each candidate detection operator runs separately in its own independent thread, and transmits information through a message queue.

[0033] It is understandable that there are multiple candidate detection operators that can perform target detection. All candidate detection operators can output target detection results belonging to the same type of target objects. The present application does not limit whether the target detection results output by each candidate detection operator are the same target area of ​​the same type of target object, as long as they are target detection results belonging to the same type of target object. The target area can be a preset area such as the facial area, body area, tail area, etc. of the target object. The target area can also be a custom area. The present application takes the target detection results output by each candidate detection operator as the same target area of ​​the same type of target object as an example. Exemplarily, when the target object of the same type is a cat, each different candidate detection operator will output target detection results for the facial area of ​​the cat belonging to different scene types.

[0034] In some application scenarios, the candidate scene type matched by the target detection operator is the same as the target scene type. In other application scenarios, the candidate scene type matched by the target detection operator is different from the target scene type. It is understandable that the candidate detection operator that matches the target scene type may be the same as the target scene type, or the candidate scene type matched by the candidate detection operator may be different from the target scene type.

[0035] Exemplarily, when the target scene type corresponding to the image to be extracted is one of the candidate scene types matched by several candidate detection operators, the candidate scene type matched by the target detection operator is the same as the target scene type. In this case, the target detection operator may be the candidate detection operator matched by the target scene type.

[0036] Exemplarily, when the target scene type corresponding to the image to be extracted is not any one of the candidate scene types matched by several candidate detection operators, the candidate scene type matched by the target detection operator is different from the target scene type. At this time, the target detection operator may be at least one candidate detection operator. In some application scenarios, other detection operators other than several candidate detection operators in the target detection model may be used as target detection operators. Other detection operators other than several candidate detection operators in the target detection model may be obtained by training the training data corresponding to several candidate detection operators in the training target detection model. In other application scenarios, when the number of target detection operators is limited to a preset number of candidate detection operators, the value of the preset number may be 1, an intermediate value, or a value corresponding to the total number of several candidate detection operators, wherein the intermediate value may be a value between 1 and the value corresponding to the total number. Specifically, when the value of the above-mentioned preset number is 1 or an intermediate value, a preset number of candidate detection operators are selected from each candidate detection operator as the target detection operator according to the first preset rule. The first preset rule may be to randomly specify a preset number of candidate detection operators from each candidate detection operator, and the candidate detection operators randomly specified each time may be different. It may also be to specify a preset number of fixed candidate detection operators from each candidate detection operator, or it may be to take turns to use a preset number of candidate detection operators as target detection operators in a certain order. Specifically, when the value of the above preset number is a value corresponding to the total number of candidate detection operators, a preset number of candidate detection operators are selected from each candidate detection operator as the target detection operator according to the second preset rule. The second preset rule includes directly using each candidate detection operator as the target detection operator.

[0037] Step S13: Use the target detection operator to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted.

[0038] The image to be extracted is input into the target detection operator for target detection, and the target detection result corresponding to the image to be extracted is obtained, wherein the target detection result includes several candidate detection frames. Several candidate detection frames may be detection frames in the image to be extracted. In some application scenarios, the above step S11 can simultaneously obtain several images to be extracted and the target scene type corresponding to each image to be extracted. That is, all images to be extracted can be input into the target detection model in batches for target detection, and the target detection result corresponding to each image to be extracted is obtained. Exemplarily, for each batch of images to be extracted, the number of each image to be extracted is multiple. The above step S11 can obtain the current batch of images to be extracted and the target scene type of each image to be extracted. Among them, the target scene types of the current batch of images to be extracted can be the same scene type or different scene types. This application takes the case that the target scene types of each batch of images to be extracted are all different scene types. The above step S12 can be based on the target scene type corresponding to each image to be extracted, and selects the candidate detection operator matching the target scene type from the several candidate detection operators in the target detection model as the target detection operator corresponding to each image to be extracted. Exemplarily, for the current batch, the target scene types of the images to be extracted in the current batch are all different scene types, and the target scene type of each image to be extracted is one of several candidate scene types. Because the target scene types of each image to be extracted are different, the target detection operators corresponding to each image to be extracted determined in the above step S12 are all different candidate detection operators. The above step S13 can be to input each image to be extracted into the target detection model and the target detection operator corresponding to each image to be extracted. At this time, the target detection operator corresponding to each image to be extracted performs target detection on each image to be extracted in parallel, and synchronously obtains the target detection results corresponding to each image to be extracted in the current batch. It can be understood that the present application does not limit the number of images to be extracted when processing each batch. The number can be one or more. The present application only takes the input of one image to be extracted into the target detection model for target detection to obtain the target detection result as an example.

[0039] It can be considered that the target detection operator is used to detect the target of the image to be extracted belonging to a specific scene, and the quality of the target detection result obtained is high, the detection rate is high, and the false detection rate is low. In addition, when each image to be extracted is input into the target detection model in batches, the target detection processing of each image to be extracted can be performed in parallel, thereby improving the efficiency of target detection of each image to be extracted.

[0040] Step S14: extract features from a number of candidate detection frames to obtain feature extraction results of the image to be extracted.

[0041] Input several candidate detection frames into the feature extraction operator for feature extraction to obtain the feature extraction result of the image to be extracted. The feature extraction operator is used to extract features from the target detection result. The feature extraction operator can be a preset feature extraction network, and different feature extraction networks or modules are set according to the feature extraction requirements. Among them, the preset feature extraction network can be a MobileNet feature extraction network, an Inception feature extraction network, etc. The setting method of the feature extraction network is not limited here.

[0042] In some application scenarios, several candidate detection frames in the target detection result are directly input into the feature extraction operator for feature extraction to obtain the feature extraction result of the image to be extracted. In other application scenarios, step S14 includes: filtering several candidate detection frames to obtain at least some candidate detection frames, which are detection frames that meet the filtering requirements, that is, valid detection frames. Feature extraction is then performed on at least some of the candidate detection frames to obtain the feature extraction result of the image to be extracted. By improving the accuracy of the detection frame input by the feature extraction operator, the accuracy of the feature value of the output valid target is improved.

[0043] The steps of training the above-mentioned target detection model may be: first group the training image data according to the image acquisition scene, and mark each image of the same scene type in the training image data as a group to obtain image data groups corresponding to each scene type. For example, the training image data can be roughly divided into an image data group corresponding to scene type A, an image data group corresponding to scene type B, and an image data group corresponding to scene type C. Then, the image data groups corresponding to different scene types are respectively input into the target detection model for training. Specifically, the image data groups corresponding to different scene types are respectively input into the candidate detection operators matched by each scene type in the target detection model for training, so that different candidate detection operators have better target detection results for the image data corresponding to the matched scene types.

[0044] The above scheme obtains the image to be extracted and the target scene type corresponding to the image to be extracted, and uses the target detection operator matching the target scene type to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted. Compared with the prior art that uses the same target detection operator to perform target detection on the same type of target objects under different scene types, the present application can improve the accuracy of the target detection result, thereby performing feature extraction on several candidate detection boxes in the target detection result, and obtaining the feature extraction result of the image to be extracted, which can improve the accuracy of the feature extraction result of the image to be extracted.

[0045] See also Figure 2 , Figure 2 This is a schematic diagram of the process of an embodiment of the feature extraction method of the present application. Figure 2 .

[0046] In some embodiments, the scene type that matches each candidate detection operator is the candidate scene type, and the above step S12 may include the following steps: Step S21: Determine whether the target scene type is one of the candidate scene types. Step S22: In response to the target scene type being one of the candidate scene types, determine the candidate detection operator corresponding to the target scene type as the target detection operator.

[0047] For each candidate scene type, the candidate scene type is any one of the scene types matched by several candidate detection operators. That is, there is a matching relationship between each candidate scene type and each candidate detection operator. A matching relationship is established between each candidate scene type and each candidate detection operator. According to the matching relationship between each candidate scene type and each candidate detection operator, each candidate detection operator is bound to the candidate scene type it matches. The above step S21 can be to determine whether the target scene type is one of the candidate scene types matched by each candidate detection operator.

[0048] In some application scenarios, when the target scene type is any one of the candidate scene types, the candidate detection operator corresponding to the target scene type is determined as the target detection operator corresponding to the image to be extracted. Exemplarily, when the target scene type corresponding to the image to be extracted is scene type A, the target detection operator is detection operator a. The detection operator a in the target detection model is used to perform target detection on the image to be extracted, and a target detection result corresponding to the image to be extracted is obtained.

[0049] In some embodiments, the feature extraction method may further include the following steps: Step S23: In response to the target scene type not being one of the candidate scene types, taking at least one of the candidate detection operators as a target detection operator.

[0050] In other application scenarios, when the target scene type is not any one of the candidate scene types, at least one of the candidate detection operators or the other detection operators are determined as the target detection operator corresponding to the image to be extracted. Exemplarily, when the target scene type corresponding to the image to be extracted is scene type D, the target detection operator is at least one of the candidate detection operators or the other detection operators, and the target detection operator is used to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted. It can be understood that the target detection operator that needs to be input for the image to be extracted can be determined with reference to the above content when the target scene type corresponding to the image to be extracted is scene type D. The target detection operator can be at least one of the candidate detection operators or the other detection operators mentioned above, and the determination method is not repeated here. Here, the target detection operator that needs to be input for the image to be extracted is at least one of the candidate detection operators as an example, which is only for illustration, and does not limit the target detection operator from being other detection operators mentioned above.

[0051] In some embodiments, the above step S13 may include the following steps: Step S24: Input the image to be extracted into each target detection operator for target detection, and obtain the candidate detection results output by each target detection operator. Step S25: Based on the candidate detection results output by each target detection operator, obtain the target detection result corresponding to the image to be extracted.

[0052] In response to the target scene type not being one of the candidate scene types, each target detection operator may be at least one candidate detection operator in the target detection model. Each candidate detection result is obtained by inputting the image to be extracted into each target detection operator for target detection. Each candidate detection result includes a plurality of detection boxes to be selected. The plurality of detection boxes may be detection boxes in the image to be extracted.

[0053] Exemplarily, when the target detection operators determined in the above step S23 are all candidate detection operators in the target detection model, the above step S24 may be to input an image to be extracted into each candidate detection operator in the target detection model, and each candidate detection operator performs target detection processing on the same image to be extracted in parallel, and obtains the candidate detection results of the image to be extracted output by each candidate detection operator. It can be understood that each candidate detection result is used to represent the target detection result obtained by each candidate detection operator performing target detection on the same image to be extracted.

[0054] In some application scenarios, the above step S25 can directly use the candidate detection results output by each target detection operator as the target detection results corresponding to the image to be extracted. The above step S14 can be to use the several to-be-selected detection frames included in the candidate detection results output by each target detection operator as the to-be-selected data. For example, the to-be-selected data are the several to-be-selected detection frames included in all candidate detection results. The to-be-selected data is filtered to obtain at least part of the to-be-selected detection frames, which are detection frames that meet the filtering requirements, that is, valid detection frames. Then, feature extraction is performed on at least part of the to-be-selected detection frames to obtain the feature extraction results of the image to be extracted. In other application scenarios, the above step S25 can first select the candidate detection results output by each target detection operator to obtain the candidate detection results output by any target detection operator, and the candidate detection results output by any target detection operator include several to-be-selected detection frames that meet the selection requirements. The selection requirement may be to use the candidate detection result with the smallest difference between the distribution of the to-be-selected detection frame and the distribution of the preset detection frame in each candidate detection result as the target detection result corresponding to the image to be extracted. Furthermore, in some application scenarios, the above step 14 may be to directly input several to-be-selected detection frames in the target detection result corresponding to the image to be extracted into a feature extraction operator for feature extraction to obtain the feature extraction result of the image to be extracted, or to filter several to-be-selected detection frames to obtain at least some of the to-be-selected detection frames and then perform feature extraction on at least some of the to-be-selected detection frames.

[0055] In some embodiments, the above step S25 may include the following steps: First, obtain the distribution of preset detection frames corresponding to each target detection operator. Then, based on the position information of several detection frames to be selected in each candidate detection result, determine the distribution of the detection frames to be selected corresponding to each candidate detection result. Subsequently, for each target detection operator, the difference between the distribution of the detection frames to be selected and the distribution of the preset detection frames is determined as the weight of the candidate detection result. Then, the candidate detection result with the largest weight is used as the target detection result corresponding to the image to be extracted.

[0056] The candidate detection results output by each target detection operator are selected and processed as described above. The candidate detection results output by any target detection operator can be obtained by referring to the following contents:

[0057] The preset detection frame distribution can characterize the position information and / or area information of the preset detection frame of the input image in the candidate detection results that the candidate detection operator can output. Different candidate detection operators match different preset detection frame distributions. Exemplarily, for the second candidate detection operator, that is, detection operator b, the preset detection frame distribution can be that the position of the preset detection frame of the input image is concentrated in the central area compared to the input image, and the ratio between the area of ​​the preset detection frame and the area of ​​the input image can be a preset value, which can be one-half.

[0058] The distribution of the detection frames to be selected can characterize the position information and / or area information of the detection frames to be selected of the image to be extracted in the candidate detection results that can be output by the target detection operator. The position information and / or area information of the detection frames to be selected is determined by the position information of the detection frames to be selected in each candidate detection result. The position information and / or area information of the detection frames to be selected is used as the distribution of the detection frames to be selected corresponding to each candidate detection result.

[0059] The weight of the candidate detection result is used to represent the evaluation value of the candidate detection result among the candidate detection results. The larger the weight of the candidate detection result, the larger the evaluation value, and the greater the possibility that the candidate detection result is determined as the target detection result corresponding to the image to be extracted. For each target detection operator, the difference between the distribution of the selected detection box and the distribution of the preset detection box is compared to determine the size of the difference. The size of the difference is inversely proportional to the weight of the candidate detection result. The larger the size of the difference, the smaller the weight of the candidate detection result.

[0060] It can be considered that the present application does not limit whether the target scene type of the image to be extracted must be any one of several candidate scene types. When the target scene type is any one of several candidate scene types, or is not any one of several candidate scene types, the image to be extracted is input into different target detection operators, and then the candidate detection results output by the target detection operator are processed accordingly, which can realize specific target detection processing and subsequent selection processing and filtering processing, and improve the accuracy of the feature extraction results by improving the accuracy of determining the detection box that needs to be extracted.

[0061] See also Figure 3 , Figure 3 This is a schematic diagram of the process of an embodiment of the feature extraction method of the present application. Figure 3 .

[0062] In some embodiments, the above step S14 may include the following steps: Step S31: Obtain a preset filtering strategy set, the preset filtering strategy set includes multiple detection frame filtering strategies. Step S32: Select at least one detection frame filtering strategy from the multiple detection frame filtering strategies as the current strategy. Step S33: Filter several candidate detection frames according to the current strategy to obtain at least one candidate detection frame, and use at least one candidate detection frame as the detection frame to be extracted. Step S34: Perform feature extraction on the detection frame to be extracted to obtain a feature extraction result.

[0063] The preset filtering strategy set includes a variety of detection frame filtering strategies. The preset filtering strategy can be preset or customized by the user. The customized preset filtering strategy set can be a dynamic setting of the desired detection frame position information and / or area information in the strategy configuration interface, or a fine-tuning of the desired detection frame position information and / or area information in each preset filtering strategy.

[0064] In some embodiments, the multiple detection frame filtering strategies include a first detection frame filtering strategy, a second detection frame filtering strategy, a third detection frame filtering strategy and / or a fourth detection frame filtering strategy, wherein the first detection frame filtering strategy includes filtering out candidate detection frames whose center point distance is within a preset distance range to obtain at least one candidate detection frame, and the center point distance is the distance between the candidate detection frame and the center point of the image to be extracted. The second detection frame filtering strategy includes filtering out candidate detection frames whose area ranking is within a preset ranking range to obtain at least one candidate detection frame, and the area ranking is the ranking between the area sizes of each candidate detection frame. The third detection frame filtering strategy includes filtering out candidate detection frames whose intersection and union ratio is within a preset range to obtain at least one candidate detection frame, and the intersection and union ratio is the area ratio between the area where each candidate detection frame is located and several preset detection frames in the image to be extracted. The fourth detection frame filtering strategy includes filtering out candidate detection frames that are in or not in several preset areas of the image to be extracted to obtain at least one candidate detection frame.

[0065] In some application scenarios, the first detection frame filtering strategy includes filtering out candidate detection frames whose center point distances are within a preset distance range to obtain at least one candidate detection frame. The center point is the pixel point corresponding to the center position of the image to be extracted. The center point distance is the distance between the candidate detection frame and the center point of the image to be extracted. Specifically, the center point distance is the distance between the preset position of the candidate detection frame and the center point of the image to be extracted. The preset position of the candidate detection frame may be the center point corresponding to the candidate detection frame. The preset distance range may be the center point distance of the preset ranking in the center point distance sorting from small to large as the threshold distance, and the center point distance within the preset distance range is less than the threshold distance. Exemplarily, only the candidate detection frame closest to the center point of the image to be extracted is retained. That is, only the candidate detection frame with the smallest center point distance is retained, and the candidate detection frame is used as the object for subsequent feature extraction.

[0066] The preset distance range, the preset position of the candidate detection frame, and the preset ranking may be preset, or may be obtained through the above-mentioned custom processing.

[0067] In some application scenarios, the second detection frame filtering strategy includes filtering out candidate detection frames whose area rankings are within a preset ranking range to obtain at least one candidate detection frame. The area size of the candidate detection frame is used to represent the area of ​​the image to be extracted occupied by the candidate detection frame. The area ranking is the ranking between the area sizes of each candidate detection frame. The preset ranking range can be the center point distance of the preset ranking in the descending order of the area size of the candidate detection frame as the threshold distance, and the area ranking of the candidate detection frame within the preset ranking range is higher than the preset ranking. Exemplarily, the second detection frame filtering strategy can be to retain only the candidate detection frame with the largest area. That is, only the candidate detection frame with the largest area ranking is retained, and the candidate detection frame is used as the object for subsequent feature extraction. The preset ranking in the preset ranking range can be preset or obtained through the above-mentioned custom processing.

[0068] In some application scenarios, the third detection frame filtering strategy includes filtering out candidate detection frames whose intersection-over-union ratio is within a preset range to obtain at least one candidate detection frame.

[0069] Several preset detection frames are detection frames in the above-mentioned preset detection frame distribution. The preset detection frames matched by different candidate detection operators have different distributions. The intersection-and-union ratio of each candidate detection frame can be the area ratio between the region where the candidate detection frame is located and several preset detection frames in the image to be extracted. The intersection-and-union ratio between each candidate detection frame and the same preset detection frame is determined. Taking the preset detection frame as a unit, the intersection-and-union ratio between each candidate detection frame and the same preset detection frame is divided into the intersection-and-union ratio group corresponding to the preset detection frame. The candidate detection frame corresponding to the maximum intersection-and-union ratio in each intersection-and-union ratio group is used as the reserved object corresponding to the intersection-and-union ratio group. In some application scenarios, several candidate detection frames are filtered according to the third detection frame filtering strategy to obtain at least one candidate detection frame that can be the reserved object corresponding to each intersection-and-union ratio group. In other applications, a preset number of reserved objects are selected from the reserved objects corresponding to each intersection-and-union ratio group as at least one candidate detection frame. The preset number of reserved objects can be the top-ranked intersection-and-union ratios. Exemplarily, only the candidate detection frame with the largest intersection-and-union ratio is retained. The preset range can be preset or obtained through the above-mentioned custom processing. The intersection-and-union ratio being within the preset range may be that the intersection-and-union ratio is the largest in each group of intersection-and-union ratios, or may be the largest in all intersection-and-union ratios.

[0070] In some application scenarios, the fourth detection frame filtering strategy includes filtering out candidate detection frames that are in or not in several preset areas of the image to be extracted to obtain at least one candidate detection frame. Exemplarily, only the candidate detection frames in the preset area are retained. The specific method may be to determine whether the center point of each candidate detection frame is in the preset area, and only the candidate detection frame whose center point is in the preset area is retained. Wherein, each preset area is a polygonal closed area. If there are multiple preset areas, the center point of the candidate detection frame is retained as long as it is in one of the preset areas, and the candidate detection frame is used as an object for subsequent feature extraction. Exemplarily, the candidate detection frame in the preset area is filtered out. The specific method may be to determine whether the center point of each candidate detection frame is outside the preset area, and only the candidate detection frame whose center point is outside the preset area is retained. Wherein, the preset area is a polygonal closed area. If there are multiple preset areas, the center point of the candidate detection frame is retained only when it is outside all preset areas, and the candidate detection frame is used as an object for subsequent feature extraction. Several preset areas may be preset or obtained through the above-mentioned custom processing.

[0071] In some application scenarios, the current strategy may be to randomly specify at least one detection frame filtering strategy from multiple detection frame filtering strategies, or to specify at least one detection frame filtering strategy from multiple detection frame filtering strategies in turn in order, or to be a fixed combination of at least one detection frame filtering strategy from multiple detection frame filtering strategies. In other application scenarios, the current strategy may be a fixed combination of at least one detection frame filtering strategy from multiple detection frame filtering strategies determined based on the target scene type corresponding to the image to be extracted.

[0072] In some embodiments, the scene type that matches each candidate detection operator is a candidate scene type, and the above step S32 may include the following steps: obtaining a matching relationship between each candidate scene type and at least one candidate strategy, each candidate strategy is used to represent a combination of at least one detection frame filtering strategy, and the matching relationship is used to represent that each candidate scene type matches a candidate strategy. Determine whether the target scene type is one of the candidate scene types. In response to the target scene type being one of the candidate scene types, determine the candidate strategy corresponding to the target scene type as the current strategy.

[0073] The candidate strategy is used to represent a fixed combination of at least one detection frame filter strategy. For example, the candidate strategy can be a combination of the first detection frame filter strategy and the second detection frame filter strategy, or a combination of the second detection frame filter strategy, the third detection frame filter strategy, and the fourth detection frame filter strategy. This is only for illustration and does not limit the specific candidate strategy.

[0074] Establish a matching relationship between each candidate scenario type and each candidate strategy. When the target scenario type is one of the candidate scenario types, determine the candidate strategy corresponding to the target scenario type as the current strategy. When the target scenario type is not any of the candidate scenario types, directly determine each candidate strategy as the current strategy. Alternatively, randomly specify, fixedly specify, or poll any one candidate strategy from the candidate strategies as the current strategy.

[0075] In some embodiments, the feature extraction method may further include the following steps: in response to the target scene type not being one of the candidate scene types, filtering a plurality of candidate detection frames according to each candidate strategy, respectively, and obtaining at least one candidate detection frame corresponding to each candidate strategy. At least one candidate detection frame corresponding to a candidate strategy that meets the selection condition is selected as the detection frame to be extracted, the selection condition is related to whether each reference factor meets the preset reference factor, each reference factor is at least one candidate detection frame corresponding to each candidate strategy, the preset reference factor includes a first preset reference factor, a second preset reference factor, a third preset reference factor and / or a fourth preset reference factor, the first preset reference factor is related to the distance between each reference factor and the center point of the image to be extracted, the second preset reference factor is related to the size of the image to be extracted occupied by each reference factor, the third preset reference factor is related to the intersection and union ratio between the area where each reference factor is located and the plurality of preset detection frames in the image to be extracted, and the fourth preset reference factor is related to the distance between each reference factor and the reference point of the plurality of preset areas in the image to be extracted.

[0076] At least one candidate detection frame corresponding to a candidate strategy that meets the selection conditions is selected as the detection frame to be extracted. The selection conditions are related to whether each reference factor meets the preset reference factor. Each reference factor is at least one candidate detection frame corresponding to each candidate strategy.

[0077] When all candidate detection results of the image to be extracted are filtered according to all candidate strategies, a reference factor corresponding to each candidate strategy will be obtained. Each reference factor is used to represent at least one candidate detection frame filtered by each candidate strategy. It can be understood that each reference factor is a filtering result obtained by filtering any candidate detection result of the image to be extracted according to each candidate strategy. Because the target scene type does not belong to any candidate scene type, all candidate detection operators are input in parallel in the image to be extracted to obtain the candidate detection results corresponding to each candidate detection operator. Each candidate detection result includes several detection frames to be selected. For each selected detection result, the candidate detection result is filtered according to each candidate strategy respectively, and the reference factors obtained by filtering the same candidate detection result under different candidate strategies are obtained. The number of each reference factor is the product of the number of each candidate strategy and the number of candidate detection results corresponding to each candidate scene.

[0078] The process of selecting at least one candidate detection frame corresponding to a candidate strategy that meets the selection conditions as the detection frame to be extracted may be selecting at least one candidate detection frame filtered by at least part of the candidate strategies from the at least one candidate detection frame filtered by each candidate strategy as the detection frame to be extracted.

[0079] In some application scenarios, the first preset reference factor may be the same as the first detection frame filtering strategy. The first preset reference factor may be the same as the first detection frame filtering strategy, or it may be obtained by fine-tuning the first detection frame filtering strategy. Exemplarily, only the candidate detection frame with the smallest center point distance is retained in each reference factor, and the reference factor corresponding to the candidate detection frame is used as the target reference factor. Then at least one candidate detection frame in the target reference factor is used as the detection frame to be extracted. In some application scenarios, the second preset reference factor may be the same as the second detection frame filtering strategy, or it may be obtained by fine-tuning the second detection frame filtering strategy. Exemplarily, only the candidate detection frame with the largest area ranking is retained in each reference factor, and the reference factor corresponding to the candidate detection frame is used as the target reference factor. Then at least one candidate detection frame in the target reference factor is used as the detection frame to be extracted. In some application scenarios, the third preset reference factor may be the same as the third detection frame filtering strategy, or it may be obtained by fine-tuning the third detection frame filtering strategy. Exemplarily, only the candidate detection frame with the largest intersection-over-union ratio is retained in each reference factor, and the reference factor corresponding to the candidate detection frame is used as the target reference factor. Then at least one candidate detection frame in the target reference factor is used as the detection frame to be extracted. In some application scenarios, the fourth preset reference factor may be the same as the fourth detection frame filtering strategy, or may be obtained by fine-tuning the fourth detection frame filtering strategy. The fourth preset reference factor includes a plurality of preset areas in the image to be extracted and a plurality of preset areas in the fourth detection frame filtering strategy, which are the same area.

[0080] The reference points of several preset areas included in the fourth preset reference factor can be any point on the area boundary range of several preset areas in the fourth detection frame filtering strategy. The distance between each reference factor and the reference points of several preset areas in the image to be extracted can be used to indicate whether each reference factor is within any preset area or outside all preset areas. Exemplarily, only candidate detection frames that are within any preset area or outside all preset areas are retained in each reference factor, and the reference factor corresponding to the candidate detection frame is used as the target reference factor. Then at least one candidate detection frame in the target reference factor is used as the detection frame to be extracted.

[0081] Exemplarily, assuming that the number of candidate strategies is 3, the number of detection frames of at least one candidate detection frame in the reference factor filtered according to each candidate strategy is 1, that is, the number of candidate detection frames in the reference factor corresponding to each candidate strategy is 1. The number of candidate scene types is 3. At this time, the number of candidate detection results of the image to be extracted is 3, and each candidate detection result is filtered according to 3 candidate strategies respectively, and 3 detection frames filtered by the 3 candidate strategies are obtained respectively. For example, the number of detection frames to be extracted is required to be 1. At this time, according to the above selection conditions, 1 candidate detection frame is selected from the candidate detection frames corresponding to the 9 reference factors as the detection frame to be extracted.

[0082] It is understandable that the preset reference factor may be at least one, at least two, at least three or all four of the first preset reference factor, the second preset reference factor, the third preset reference factor and the fourth preset reference factor. That is, the preset reference factor may be used for a combination of the first preset reference factor, the second preset reference factor, the third preset reference factor and the fourth preset reference factor. Exemplarily, in the case where each preset reference factor is equivalent to each detection frame filtering strategy, in the case where any candidate strategy is a combination of the first detection frame filtering strategy and the second detection frame filtering strategy, the preset reference factor may be a combination of the first preset reference factor and the second preset reference factor, or may not be a combination of the first preset reference factor and the second preset reference factor. This is only an example, in the case where each preset reference factor is equivalent to each detection frame filtering strategy, the first combination method is a combination method of each detection frame filtering strategy corresponding to the candidate strategy that needs to be used in the process of obtaining the reference factor corresponding to each candidate strategy, and the second combination method is a combination method of each preset reference factor in the selection condition. It is understandable that the present application does not limit whether the first combination method is the same as the second combination method.

[0083] According to at least one detection frame filtering strategy of the current strategy, several candidate detection frames corresponding to the image to be extracted are filtered to obtain at least one candidate detection frame, and the at least one candidate detection frame is used as the detection frame to be extracted. The detection frame to be extracted is used as an object for subsequent feature extraction.

[0084] It can be considered that the above steps S31, S32 and S33 can be executed by a filtering module in the feature extraction device. The filtering module can have a first function and / or a second function. The first function can be to summarize the target detection results corresponding to multiple candidate detection operators to ensure that when feature extraction is performed on several candidate detection frames in each picture to be extracted, the several candidate detection frames are target detection results output by only any one candidate detection operator. The second function includes performing custom filtering on several candidate detections of the target detection of each picture to be extracted according to the registered custom target filtering callback function, retaining valid targets, and the retained valid targets are the detection frames to be extracted so that feature extraction is only performed on the detection frames to be extracted in the future, thereby improving the efficiency and accuracy of feature extraction.

[0085] The feature extraction of the detection frame to be extracted can be performed in a manner that the feature extraction result is obtained by extracting features of the region where the detection frame to be extracted is located in the image to be extracted. The image to be extracted is cropped according to the region where the detection frame to be extracted is located to obtain a cropped image. The cropped image is input into the above-mentioned feature extraction operator to obtain the feature value of the effective target.

[0086] The above scheme obtains the image to be extracted and the target scene type corresponding to the image to be extracted, and uses the target detection operator matching the target scene type to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted. Compared with the prior art that uses the same target detection operator to perform target detection on the same type of target objects under different scene types, the present application can improve the accuracy of the target detection result, thereby performing feature extraction on several candidate detection boxes in the target detection result, and obtaining the feature extraction result of the image to be extracted, which can improve the accuracy of the feature extraction result of the image to be extracted.

[0087] See also Figure 4 , Figure 4It is a structural schematic diagram of an embodiment of the feature extraction device of the present application. The feature extraction device 40 includes an acquisition module 41, a selection module 42, a target detection module 43 and a feature extraction module 44; the acquisition module 41 is used to acquire the image to be extracted and the target scene type corresponding to the image to be extracted, and the target scene type is used to represent the scene type of the image acquisition scene where the image to be extracted is located; the selection module 42 is used to select a candidate detection operator that matches the target scene type from a number of candidate detection operators in the target detection model as a target detection operator, and different candidate detection operators can perform target detection on the same type of target objects under different scene types; the target detection module 43 is used to perform target detection on the image to be extracted using the target detection operator, and obtain the target detection result corresponding to the image to be extracted, and the target detection result includes a number of candidate detection frames; the feature extraction module 44 is used to perform feature extraction on a number of candidate detection frames, and obtain the feature extraction result of the image to be extracted.

[0088] The above scheme obtains the image to be extracted and the target scene type corresponding to the image to be extracted, and uses the target detection operator matching the target scene type to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted. Compared with the prior art that uses the same target detection operator to perform target detection on the same type of target objects under different scene types, the present application can improve the accuracy of the target detection result, thereby performing feature extraction on several candidate detection boxes in the target detection result, and obtaining the feature extraction result of the image to be extracted, which can improve the accuracy of the feature extraction result of the image to be extracted.

[0089] Please refer to the feature extraction method for the functions performed by each module, which will not be repeated here.

[0090] See also Figure 5 , Figure 5 1 is a schematic diagram of the structure of an embodiment of an electronic device of the present application. The electronic device 50 includes a memory 51 and a processor 52, and the processor 52 is used to execute program instructions stored in the memory 51 to implement the steps in the above-mentioned feature extraction method embodiment. In a specific implementation scenario, the electronic device 50 may include but is not limited to: a multi-camera device, a microcomputer, and a server. In addition, the electronic device 50 may also include a mobile device such as a laptop computer and a tablet computer, which is not limited here.

[0091] Specifically, the processor 52 is used to control itself and the memory 51 to implement the steps in the above-mentioned feature extraction method embodiment. The processor 52 can also be called a CPU (Central Processing Unit). The processor 52 may be an integrated circuit chip with signal processing capabilities. The processor 52 can also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 52 can be implemented by an integrated circuit chip.

[0092] The above scheme obtains the image to be extracted and the target scene type corresponding to the image to be extracted, and uses the target detection operator matching the target scene type to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted. Compared with the prior art that uses the same target detection operator to perform target detection on the same type of target objects under different scene types, the present application can improve the accuracy of the target detection result, thereby performing feature extraction on several candidate detection boxes in the target detection result, and obtaining the feature extraction result of the image to be extracted, which can improve the accuracy of the feature extraction result of the image to be extracted.

[0093] See also Figure 6 , Figure 6 The computer-readable storage medium 60 stores program instructions 601, which implement the steps in any of the above-mentioned feature extraction method embodiments when executed by a processor.

[0094] The above scheme obtains the image to be extracted and the target scene type corresponding to the image to be extracted, and uses the target detection operator matching the target scene type to perform target detection on the image to be extracted to obtain the target detection result corresponding to the image to be extracted. Compared with the prior art that uses the same target detection operator to perform target detection on the same type of target objects under different scene types, the present application can improve the accuracy of the target detection result, thereby performing feature extraction on several candidate detection boxes in the target detection result, and obtaining the feature extraction result of the image to be extracted, which can improve the accuracy of the feature extraction result of the image to be extracted.

[0095] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0096] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0097] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0098] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0099] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. 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 medium, including a number of instructions 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 application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

Claims

1. A feature extraction method, characterized in that: The method comprises: Acquire an image to be extracted and a target scene type corresponding to the image to be extracted, wherein the target scene type is used to indicate the scene type of an image acquisition scene in which the image to be extracted is located; Selecting a candidate detection operator that matches the target scene type from a plurality of candidate detection operators in the target detection model as the target detection operator, wherein different candidate detection operators can perform target detection on the same type of target objects in different scene types; Performing target detection on the image to be extracted by using the target detection operator to obtain a target detection result corresponding to the image to be extracted, wherein the target detection result includes a plurality of candidate detection boxes; Feature extraction is performed on the plurality of candidate detection frames to obtain a feature extraction result of the image to be extracted.

2. The method according to claim 1, characterized in that The step of performing feature extraction on the plurality of candidate detection frames to obtain a feature extraction result of the image to be extracted includes: Obtaining a preset filtering strategy set, wherein the preset filtering strategy set includes multiple detection frame filtering strategies; Selecting at least one detection frame filtering strategy from the multiple detection frame filtering strategies as a current strategy; Filtering the plurality of candidate detection frames according to the current strategy to obtain at least one candidate detection frame, and using the at least one candidate detection frame as the detection frame to be extracted; Perform feature extraction on the detection frame to be extracted to obtain the feature extraction result.

3. The method according to claim 2, characterized in that The scene type that matches each of the candidate detection operators is a candidate scene type, and the selecting at least one detection frame filtering strategy from the multiple detection frame filtering strategies as the current strategy includes: Acquire a matching relationship between each of the candidate scene types and at least one candidate strategy, where each of the candidate strategies is used to represent a combination of the at least one detection frame filtering strategy, and the matching relationship is used to represent that each of the candidate scene types matches one candidate strategy; Determining whether the target scene type is one of the candidate scene types; In response to the target scene type being one of the candidate scene types, the candidate policy corresponding to the target scene type is determined as the current policy.

4. The method according to claim 3, characterized in that The method further comprises: In response to the target scene type not being one of the candidate scene types, filtering the plurality of candidate detection frames according to the candidate strategies respectively to obtain at least one candidate detection frame corresponding to each of the candidate strategies respectively; At least one candidate detection frame corresponding to the candidate strategy that meets the selection condition is selected as the detection frame to be extracted, wherein the selection condition is related to whether each reference factor meets a preset reference factor, and each reference factor is at least one candidate detection frame corresponding to each candidate strategy. The preset reference factors include a first preset reference factor, a second preset reference factor, a third preset reference factor and / or a fourth preset reference factor. The first preset reference factor is related to the distance between each of the reference factors and the center point of the image to be extracted. The second preset reference factor is related to the size of the image to be extracted occupied by each of the reference factors. The third preset reference factor is related to the intersection-and-union ratio between the area where each of the reference factors is located and several preset detection frames in the image to be extracted. The fourth preset reference factor is related to the distance between each of the reference factors and the benchmark points of several preset areas in the image to be extracted.

5. The method according to claim 2, characterized in that: The multiple detection frame filtering strategies include a first detection frame filtering strategy, a second detection frame filtering strategy, a third detection frame filtering strategy and / or a fourth detection frame filtering strategy, The first detection frame filtering strategy includes filtering out candidate detection frames whose center point distance is within a preset distance range to obtain the at least one candidate detection frame, and the center point distance is the distance between the candidate detection frame and the center point of the image to be extracted; The second detection frame filtering strategy includes filtering out candidate detection frames whose area rankings are within a preset ranking range to obtain the at least one candidate detection frame, wherein the area ranking is a ranking between the area sizes of the candidate detection frames; The third detection frame filtering strategy includes filtering out candidate detection frames whose intersection-over-union ratio is within a preset range to obtain the at least one candidate detection frame, wherein the intersection-over-union ratio is an area ratio between an area where each candidate detection frame is located and a plurality of preset detection frames in the image to be extracted; The fourth detection frame filtering strategy includes filtering out candidate detection frames that are in or not in several preset areas of the image to be extracted to obtain the at least one candidate detection frame.

6. The method according to any one of claims 1 to 5, characterized in that The scene type that matches each of the candidate detection operators is a candidate scene type, and the step of selecting a candidate detection operator that matches the target scene type from a plurality of candidate detection operators in the target detection model as a target detection operator includes: Determining whether the target scene type is one of the candidate scene types; In response to the target scene type being one of the candidate scene types, a candidate detection operator corresponding to the target scene type is determined as the target detection operator.

7. The method according to claim 6, characterized in that The method further comprises: In response to the target scene type not being one of the candidate scene types, using at least one of the candidate detection operators as the target detection operator; The using the target detection operator to perform target detection on the image to be extracted to obtain a target detection result corresponding to the image to be extracted includes: Inputting the image to be extracted into each of the target detection operators to perform target detection, and obtaining candidate detection results output by each of the target detection operators; Based on the candidate detection results output by each of the target detection operators, the target detection result corresponding to the image to be extracted is obtained.

8. The method according to claim 7, characterized in that The step of obtaining the target detection result corresponding to the image to be extracted based on the candidate detection results output by each target detection operator includes: Obtaining the distribution of preset detection boxes corresponding to each of the target detection operators; Determine the distribution of the to-be-selected detection frames corresponding to each of the candidate detection results based on the position information of the to-be-selected detection frames in each of the candidate detection results; For each of the target detection operators, determining the difference between the distribution of the to-be-selected detection box and the distribution of the preset detection box as the weight of the candidate detection result; The candidate detection result with the largest weight is used as the target detection result corresponding to the image to be extracted.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores program instructions, and the processor retrieves the program instructions from the memory to execute the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, they are used to implement the method according to any one of claims 1 to 8.

Images