A target detection network position classification method and device and electronic equipment
By dividing the parameters of the target detection network into classification vectors and using the response values to determine the target box parameters, the error problem of the target detection network under low-bit quantization is solved, the accuracy is improved and the calculation is simplified.
Patent Information
- Application Number
- CN202111136453.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-27
- Publication Date
- 2025-10-17
- Estimated Expiration
- 2041-09-27
AI Technical Summary
The existing coordinate regression method in the target detection network under low-bit quantization leads to large errors and significant loss of accuracy.
The parameters used to determine the target frame in the target detection network are divided into classification vectors according to a preset length. The target frame parameters are determined by obtaining the value with the largest response value in the classification vector to avoid errors caused by nonlinear transformation.
Under the premise of ensuring the full-precision network effect, the accuracy of the overall network after quantization is greatly improved and the calculation process is simplified.
Smart Images
Figure CN113989521B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target detection using a neural network algorithm, and in particular to a position classification method, device, and electronic equipment for a target detection network. Background Art
[0002] Object detection is a key research area in the field of neural networks, with widespread application across various industries. However, the more effective a neural network is, the more complex the model becomes, which poses significant challenges for mobile deployment. Consequently, numerous neural network model compression methods have emerged, with quantization being a common approach. However, neural networks for different tasks have their own unique characteristics, and current quantization methods are often designed for experiments on classification networks. Even if they achieve good results, they are limited to classification networks.
[0003] Object detection tasks are more complex than classification tasks, so applying the same quantization method directly to an object detection network will result in significant accuracy loss. This is especially true when the entire network (including the input and output layers) is quantized.
[0004] At present, target detection often uses regression methods to predict coordinate frames, such as Figure 1 As shown, the input features Figure 1 After the neural network 2 is used to calculate the coordinate offset value 3 (tx, ty, tw, th), where tx and ty are the horizontal and vertical offset values of the target box center coordinate, respectively, and tw and th are the length and width offset values of the target box, respectively. After (tx, ty, tw, th) is subjected to nonlinear mathematical transformation and scaling and translation in subsequent processing, the true target box parameters can be obtained. However, when the network is low-bit quantized, the nonlinear transformation of the target box size will amplify the error. Summary of the Invention
[0005] The present invention provides a position classification method, device and electronic device for a target detection network, which are used to solve the defect of the existing technology that the coordinate frame is predicted using a regression method under low-bit quantization, and realize the coordinate regression part of the target detection network to be implemented using a classification method. Under the premise of ensuring the full-precision network effect, the accuracy of the overall network after quantization is greatly improved.
[0006] The present invention provides a location classification method for a target detection network, comprising:
[0007] Divide the parameters used to determine the target box in the target detection network into classification vectors according to the preset length;
[0008] The position parameter corresponding to the maximum response value in the classification vector or the position parameter after mathematical transformation is determined as the target frame parameter of the target detection network.
[0009] According to the position classification method of the target detection network provided by the application, the parameters for determining the target frame are specifically as follows:
[0010] The length and width of the target frame.
[0011] According to the position classification method of the target detection network provided by the application, the preset length is preset based on the size of the input feature map of the target detection network.
[0012] According to the position classification method of the target detection network provided by the application, after the position parameter corresponding to the maximum response value in the classification vector or the position parameter after mathematical transformation is determined as the target frame parameter of the target detection network, the confidence value of the target detection network is further predicted.
[0013] The numerical value between 0 and 1 is divided into a plurality of confidence vectors according to a preset difference value.
[0014] The response value of the confidence vector is obtained, and the maximum value or the value after mathematical transformation of the maximum value in the plurality of confidence vectors is taken as the predicted confidence value of the target detection network.
[0015] The application further provides a position classification device of a target detection network, comprising:
[0016] The classification module is used for dividing the parameters for determining the target frame in the target detection network into classification vectors according to a preset length.
[0017] The processing module is used for obtaining the response value of the classification vector, and determining the position parameter corresponding to the maximum response value in the classification vector or the position parameter after mathematical transformation as the target frame parameter of the target detection network.
[0018] According to the position classification device of the target detection network provided by the application, the classification module is used for dividing the parameters for determining the target frame in the target detection network into classification vectors according to a preset length.
[0019] According to the position classification device of the target detection network provided by the application, the classification module is further used for dividing the numerical value between 0 and 1 into a plurality of confidence vectors according to a preset difference value.
[0020] The processing module is further configured to obtain response values of the confidence vectors, and take a value with the largest response value in the confidence vectors or a value obtained by mathematical transformation of the largest value as a predicted confidence value of the target detection network.
[0021] The application further provides an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the position classification method of the target detection network according to any one of the above embodiments when executing the program.
[0022] The application further provides a non-transitory computer readable storage medium, which stores a computer program, wherein the computer program implements the steps of the position classification method of the target detection network according to any one of the above embodiments when executed by a processor.
[0023] The application further provides a computer program product, which includes a computer program, wherein the computer program implements the steps of the position classification method of the target detection network according to any one of the above embodiments when executed by a processor.
[0024] The application provides a position classification method, device and electronic device of a target detection network, which replaces the coordinate regression part of the target detection network with a classification method, i.e., dividing parameters obtained by the target detection network for determining a target frame into classification vectors according to a preset length, and determining the target frame by obtaining a position corresponding to a maximum response value in the classification vectors, thereby avoiding the disadvantage of amplifying errors caused by nonlinear transformation of the parameters when the coordinate regression method is used, greatly improving the precision of the whole network after quantization on the premise of ensuring the full-precision network effect. Meanwhile, the preset length can be flexibly set according to actual needs, thereby greatly simplifying the calculation process on the basis of meeting the prediction precision requirement of the target frame. BRIEF DESCRIPTION OF DRAWINGS
[0025] In order to more clearly illustrate the technical solutions of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.
[0026] Figure 1 is a schematic diagram of predicting a target frame by using a regression method in target detection;
[0027] Figure 2 is a flowchart of the position classification method of the target detection network provided by the present application;
[0028] Figure 3 is a plan view of the network output feature map of the target detection network.
[0029] Figure 4 is a schematic diagram of a target detection network position classification method provided by the application for predicting a target frame;
[0030] Figure 5 is a flowchart of a confidence value prediction classification method of a target detection network provided by the application;
[0031] Figure 6 is a structural schematic diagram of a target detection network position classification device provided by the application;
[0032] Figure 7 is a structural schematic diagram of an electronic device provided by the application. DETAILED DESCRIPTION
[0033] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0034] The present application will be described below in combination with Figures 2 to 5 A target detection network position classification method is described in the present application, and the method comprises the following steps:
[0035] 201. Parameters for determining a target frame in a target detection network are divided into classification vectors according to a preset length;
[0036] 202. Response values of the classification vectors are obtained, and a position parameter corresponding to a maximum value of the response values in the classification vectors or a position parameter after mathematical transformation of the position parameter is determined as a target frame parameter of the target detection network.
[0037] Specifically, in target detection, the real position of a target frame needs to be determined according to the coordinates of the landmark points of the target frame or the length and width of the target frame, and the landmark points can be freely selected according to requirements, for example, the landmark points can be the center point or a corner point of the target frame, and then the length and width of the target frame can be used to determine the position of the target frame; the landmark points can also be two opposite corner points of the target frame, and the position of the target frame can also be determined through the two opposite corner points. Therefore, as long as the relative relationship between the length and width of the target frame, the corner point or the center point of the target frame and the landmark points is predetermined, the real position of the target frame can be determined through the coordinates of the landmark points or the length and width of the target frame.
[0038] Therefore, in an embodiment of the present application, the parameters for determining the target frame are set to include the coordinates of the landmark points of the target frame or the length and width of the target frame.
[0039] Specifically, the method predicts the target frame with a classification idea, i.e., setting a classification vector for the landmark point coordinates or the length and width of the target frame, and determining the target frame by obtaining the position corresponding to the maximum response value in the classification vector. More specifically, as shown in Figure 4 , the input features Figure 1 After passing through the neural network 2, the output related parameters 4 of the target frame are ((x0, x1, x2, …, xn), (y0, y1, y2, …, ym), (w0, w1, w2, …, wp), (h0, h1, h2, …, hk)), where the four groups of vectors x, y, w, and h are classification vectors of the horizontal and vertical coordinates of the center point of the target frame and classification vectors of the length and width of the target frame. Assuming that X, Y, W, and H are the horizontal and vertical coordinate parameters of the center point of the target frame, then X = argmax(x0, x1, x2, …, xn), Y = argmax(y0, y1, y2, …, ym), W = argmax(w0, w1, w2, …, wp), and H = argmax(h0, h1, h2, …, hk), where argmax() means taking the position corresponding to the maximum value in the vector, and then X, Y, W, and H are the parameters of the target frame. Further, according to different classification vector setting methods, X, Y, W, and H can be further transformed into the final target frame parameters through simple scaling or translation.
[0040] Further, taking a target detection network with an input feature map size of 416*416 and an output feature map size of 13*13 as an example, and taking the center point of the target frame as the landmark point. After the operation of the intermediate network layer, the length and width of the input feature map are reduced by 32 times, becoming the output of the network. Since the feature map size is reduced by 32 times, one output point can represent the length of 32 input points. Assuming that the input coordinates take the top-left corner point as the origin, then the point in the first row and the first column of the output represents the region with horizontal and vertical coordinate ranges of 0-31 and 0-31 in the original image, the point in the first row and the second column of the output represents the region with horizontal and vertical coordinate ranges of 32-63 and 0-31 in the original image, and the point in the nth row and the mth column of the output represents the region with horizontal and vertical coordinate ranges of (m-1)*32-m*32-1 and (n-1)*32-n*32-1 in the original image. The plan view of the network output feature map of the target detection network is shown in Figure 3 , the output feature map size is 13*13, Cx and Cy are the reduction multiples of the horizontal and vertical directions from the input image to the convolution result, i.e., each grid point in Figure 3 corresponds to a region of Cx*Cy in the input feature map.
[0041] Because the output size is determined, the Figure 3Each grid point in the output feature map represents a range in the original image, and since each grid point is responsible for detecting the target frame in which the mark point falls, it is only necessary to classify the coordinates of the mark point in each small grid point and add the coordinates of the grid point in the input feature map. Assuming that the coordinates of the mark point are relative to the coordinates of the upper left corner of the small grid point, the coordinates of each grid point are relative to the coordinates of the upper left corner of the input feature map. Taking the mark point with coordinates (9, 8) detected by the grid point in the 4th row and the 5th column as an example, since the size of the output feature map is determined, the position of each grid point in the original image is determined, and the coordinates of the upper left corner of the input feature map corresponding to the grid point in the 4th row and the 5th column can be calculated by the above n, m related formula ((5-1)*32, (4-1)*32), that is (128, 96), and the position of the mark point in the input feature map can be determined by adding the offset of the mark point relative to the grid point (9, 8) (128+9, 96+8), that is (137, 104). Next, the predicted length of each grid point is classified, and assuming that the position of the maximum response value in the length classification vector is 90 and the position of the maximum response value in the width classification vector is 80, the target frame predicted by the grid point is a rectangular frame with the center point (137, 104), the length 90 and the width 80. The above is an example of the process of determining the specific position of the target frame by the position classification method of the target detection network.
[0042] It can be understood that the present application divides the coordinates or length and width of the target frame obtained by the target detection network into classification vectors according to the preset length, determines the target frame by obtaining the position corresponding to the maximum response value in the classification vector, avoids the disadvantage of amplifying the error caused by the nonlinear transformation of parameters when the coordinate regression method is used, and greatly improves the precision of the whole network after quantization on the premise of ensuring the full-precision network effect.
[0043] In another embodiment of the present application, the preset length is preset based on the size of the input feature map of the target detection network.
[0044] Specifically, still taking the above-mentioned target detection network with the input feature map size of 416*416 and the output feature map size of 13*13 as an example, for the classification of the coordinates of the mark point, the length and width of the input feature map are W=416 and H=416 respectively, so Cx=W / 13=32 and Cy=H / 13=32. Since each grid point represents a region of 0-32 in the horizontal direction, and the mark point is in this region, when the preset length step of the horizontal coordinate of the mark point is 1, the horizontal coordinate of the mark point can be divided into 33 categories, each category representing a length, such as: the first category represents that the horizontal coordinate of the mark point is 0, the second category represents that the horizontal coordinate of the mark point is 1, the third category represents that the horizontal coordinate of the mark point is 2, and the kth category represents that the horizontal coordinate of the mark point is k-1. The vertical coordinate classification design is the same.
[0045] It is understandable that when there are higher requirements for detection accuracy, the classification can be refined. For example, if the classification is divided into 65 categories, the first category represents the horizontal coordinate of the marker point is 0, the second category represents the horizontal coordinate of the marker point is 0.5, the third category represents the horizontal coordinate of the marker point is 1, and the kth category represents the horizontal coordinate of the marker point is (k-1)*0.5. Similarly, when the accuracy requirement is not high, the classification can also be coarsened.
[0046] Furthermore, regarding the classification of target frame length, the target frame is internal to the image. Therefore, the length and width of the target frame cannot exceed the image size. Since the length and width of the input image are both 416, the length and width of the target frame range from 0 to 416. When the preset length step of the target frame length is set to 1, the target frame length categories can be directly set to 417. That is, the first category represents a detection frame length of 0, the second category represents a detection frame length of 1, the third category represents a detection frame length of 2, and the kth category represents a detection frame length of k-1. The same classification can be applied to the target frame width. Similarly, as with the coordinate classification of the landmark points, the classification can be refined or coarsened based on the accuracy requirements. For example, with double-coarsening of the length and width classification, the target frame length can be (416 / 2+1) = 209 categories. Then, the first category represents a detection frame length of 0*2, the second category represents a detection frame length of 1*2, the third category represents a detection frame length of 2*2, and the kth category represents a detection frame length of (k-1)*2.
[0047] It can be seen that the category setting of the coordinates and length and width of the marker points is based on the length and width of the input feature map, and the category is set by refining or coarsening on this basis. In addition to the above-mentioned methods that conform to certain rules, the category refinement and coarsening methods can also be set according to actual needs, that is, the preset length can be flexibly set according to actual needs, and in the position classification method of the present invention, it is only necessary to clarify the length corresponding to each category. This makes it possible to greatly simplify the calculation process when using the method of the present invention to determine the position of the target frame of the target detection network while meeting the prediction accuracy requirements of the target frame.
[0048] Further, such as Figure 5 As shown, in another embodiment of the present invention, after determining the position parameter corresponding to the value with the largest response value in the classification vector or the position parameter after mathematical transformation as the target frame parameter of the target detection network, it also includes predicting the confidence value of the target detection network:
[0049] 501. Divide the values between 0 and 1 into multiple confidence vectors according to preset differences;
[0050] 502、acquiring the response value of the confidence vector, and taking the value with the largest response value in the plurality of confidence vectors or the value after mathematical transformation of the largest value as the predicted confidence value of the target detection network.
[0051] Specifically, the confidence is a value between 0 and 1, by setting a preset difference value, the confidence can be divided into several parts, that is, the confidence is given an exact value, for example, if the preset difference value is 0.1, the confidence value is divided into: 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1, and the classification is realized by dividing into 11 confidence vectors, and then the value with the largest response value in the confidence vector is obtained by calculation, that is, the predicted confidence value of the target detection network.
[0052] It can be understood that as long as the target detection network involving coordinate regression can be realized by using the position classification method of the target detection network.
[0053] For example, case 1: taking the center point of the target frame as the mark point, combining the long and wide positioning of the target frame, taking the yolo target detection network with the input feature map size of 416*416, the output feature map size of 13*13, and the detection of 20 (hereinafter represented by cls, that is, cls=20) kinds of objects as an example, the target detection network detects the target frame under three kinds of preset anchor points, and needs a mark point and the length and width of the detection frame when positioning the target frame, wherein the horizontal coordinate of the mark point of the target frame detected by the nth row mth column of the 13*13 output should be X=m*Cx+(k-1), and the vertical coordinate is Y=n*Cy+(k-1), when the position classification method of the target detection network is adopted, the horizontal and vertical coordinates of the mark point of the target frame are classified by the length of each grid point as the preset length, and the length and width of the target frame are classified by the 2 times of the unit length of the size of the input image as the preset length, then the setting category of the horizontal and vertical coordinates of the mark point of the target frame is 33 categories from 0 to 32, the setting category of the length or width of the target frame is 209 categories from 0 to 208, and each point on the plane of the output feature map has three kinds of anchor point frames, each kind of anchor point frame is responsible for predicting a target frame, then the information of each target frame is contained in a vector with a length of 1+cls+W / 2+1+H / 2+1+Cx+1+Cy+1, and the vector length can be calculated according to the above conditions as 505. Taking the point corresponding to the second row and the third column of the output feature map of the first kind of anchor point frame as an example, the prediction result vector corresponding to the point is output, the confidence value is conf=output[0], the object category vector is objcls=argmax(output[1:21]), the length vector of the target frame is clsw=argmax(output[21:230]), the width vector of the target frame is clsh=argmax(output[230:439]), the horizontal coordinate vector of the mark point of the target frame is clsCx=argmax(output[439:472]), and the vertical coordinate vector of the mark point of the target frame is clsCy=argmax(output[472:]), so the prediction result of the point in the second row and the third column under the first kind of anchor is the confidence conf, the object category is the first objcls, the length of the target frame is (clsw-1)*2, the width is (clsh-1)*2, the horizontal coordinate of the center point is 3*32+(clsCx-1), and the vertical coordinate of the center point is 2*32+(clsCy-1).
[0054] Case 2, the target frame has been initially positioned, and the frame needs to be offset to obtain a more accurate target frame. Taking FasterRCNN as an example, the target detection network first obtains a batch of detection frames in the image to be detected through the RPN method, then filters out a batch of detection frames that may contain objects through a series of rules, then calculates the landmark point coordinates and the length and width offset of the frame through the regression method, and finally obtains the final target frame through mathematical transformation. The selected landmark point is the center point of the target frame. Since the elements to be classified are similar to those in case 1, the same classification idea as in case 1 can be used.
[0055] Case 3, the specific position of the target frame is determined by finding the diagonal points of the target frame. For the target detection network that regresses the diagonal point coordinates to determine the detection frame, there are two landmark points, which are a pair of diagonal points of the target frame. If the regression is converted to classification of the coordinates of the diagonal points, then the classification of the landmark point coordinates is set based on the length and width dimensions of the input image, and according to the actual needs, it is determined whether to refine the classification or to coarsen the classification, and then the classification of each landmark point coordinate is determined.
[0056] The following will be described in combination with Figure 6 A target detection network position classification device provided by the application is described, and the target detection network position classification device described below can be correspondingly referred to the target detection network position classification method described above.
[0057] As Figure 6 shown, the target detection network position classification device provided by the application includes a classification module 610 and a processing module 620. Among them,
[0058] The classification module 610 is used to divide the parameters for determining the target frame in the target detection network into classification vectors according to a preset length;
[0059] The processing module 620 is used to obtain the response value of the classification vector, and determine the position parameter corresponding to the value with the largest response value in the classification vector or the position parameter after mathematical transformation of the position parameter as the target frame parameter of the target detection network.
[0060] In one preferred embodiment, the classification module 610 is used to divide the landmark point coordinates for determining the target frame or the length and width of the target frame in the target detection network into classification vectors according to a preset length.
[0061] In another preferred embodiment, the classification module 610 is used to divide the landmark point coordinates of the target frame or the length and width of the target frame obtained by the target detection network into classification vectors according to a preset length preset based on the size of the input feature map of the target detection network.
[0062] In another preferred solution, the classification module 610 is further configured to divide the numerical value between 0 and 1 into a plurality of confidence vectors according to a preset difference value;
[0063] The processing module 620 is further configured to obtain a response value of the confidence vector, and take the maximum value of the response values of the plurality of confidence vectors or a value obtained by mathematical transformation of the maximum value as a predicted confidence value of the target detection network.
[0064] Specifically, the device predicts the target frame by using the classification idea, that is, sets a classification vector for the landmark point coordinates or the length and width of the target frame, determines the target frame by obtaining the position corresponding to the maximum response value in the classification vector, avoids the disadvantage of amplifying the error due to the nonlinear transformation of the parameters when the coordinate regression method is used, greatly improves the precision of the quantized network as a whole under the premise of ensuring the full-precision network effect, and simultaneously, the preset length can be flexibly set according to actual needs, greatly simplifies the calculation process on the basis of meeting the prediction precision requirement of the target frame.
[0065] The target detection network position classification device provided in the embodiment of the application is used for signing the target detection network position classification method of each embodiment. Each module included in the target detection network position classification device implements the specific method and process of the corresponding function, and the details are described in the above embodiment of the target detection network position classification method, which will not be repeated here.
[0066] The target detection network position classification device of the application is used for the target detection network position classification method of each embodiment described above. Therefore, the description and definition in the target detection network position classification method in each embodiment described above can be used to understand each execution module in the embodiment of the application.
[0067] Figure 7 An example of an entity structure schematic diagram of an electronic device is shown in Figure 7 As shown, the electronic device can include a processor 710, a communications interface 720, a memory 730 and a communications bus 740, wherein the processor 710, the communications interface 720 and the memory 730 complete mutual communication through the communications bus 740. The processor 710 can invoke the logic instructions in the memory 730 to execute the target detection network position classification method, which includes:
[0068] 201. dividing parameters for determining a target frame in a target detection network into classification vectors according to a preset length;
[0069] 202、obtaining a response value of the classification vector, and determining the position parameter corresponding to the maximum value in the classification vector or the position parameter after mathematical transformation as the target frame parameter of the target detection network.
[0070] In addition, the logic instructions in the memory 730 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.
[0071] In another aspect, the present application also provides a computer program product, which includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the position classification method of the target detection network provided by the above-mentioned methods, and the method includes:
[0072] 201、dividing the parameters for determining the target frame in the target detection network into a classification vector according to a preset length;
[0073] 202、obtaining a response value of the classification vector, and determining the position parameter corresponding to the maximum value in the classification vector or the position parameter after mathematical transformation as the target frame parameter of the target detection network.
[0074] In another aspect, the present application also provides a computer program product, which includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the position classification method of the target detection network provided by the above-mentioned methods, and the method includes:
[0075] 201、dividing the parameters for determining the target frame in the target detection network into a classification vector according to a preset length;
[0076] 202、obtaining a response value of the classification vector, and determining the position parameter corresponding to the maximum value in the classification vector or the position parameter after mathematical transformation as the target frame parameter of the target detection network.
[0077] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.
[0078] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.
[0079] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A location classification method for a target detection network, characterized in that: include: Divide the parameters used to determine the target box in the target detection network into classification vectors according to the preset length; Obtaining a response value of the classification vector, and determining a position parameter corresponding to a value with the largest response value in the classification vector or a position parameter after mathematical transformation of the position parameter as a target frame parameter of the target detection network; The parameters for determining the target frame are set to include: coordinates of marker points of the target frame or the length and width of the target frame.
2. The position classification method of the target detection network according to claim 1, characterized in that The parameters used to determine the target frame specifically include: The coordinates of the target frame's landmarks or the length and width of the target frame.
3. The position classification method of the target detection network according to claim 2, characterized in that The preset length is preset based on the size of the input feature map of the target detection network.
4. The position classification method of the target detection network according to claim 1, characterized in that After determining the position parameter corresponding to the value with the largest response value in the classification vector or the position parameter after mathematical transformation as the target frame parameter of the target detection network, the method further includes predicting the confidence value of the target detection network: Divide the values between 0 and 1 into multiple confidence vectors according to the preset difference; Obtain a response value of the confidence vector, and use a value with the largest response value among the multiple confidence vectors or a value after mathematical transformation of the largest value as the predicted confidence value of the target detection network.
5. A position classification device for a target detection network, characterized in that: include: A classification module is used to divide the parameters used to determine the target box in the target detection network into classification vectors according to a preset length; a processing module, configured to obtain a response value of the classification vector, and determine a position parameter corresponding to a value with the largest response value in the classification vector or a position parameter after mathematical transformation of the position parameter as a target frame parameter of the target detection network; The parameters for determining the target frame are set to include: coordinates of marker points of the target frame or the length and width of the target frame.
6. The position classification device of the target detection network according to claim 5, characterized in that The classification module is used to divide the coordinates of the landmark points used to determine the target frame in the target detection network or the length and width of the target frame into classification vectors according to a preset length.
7. The position classification device of the target detection network according to claim 5, characterized in that The classification module is further configured to divide the values between 0 and 1 into a plurality of confidence vectors according to a preset difference; The processing module is also used to obtain the response value of the confidence vector, and use the value with the largest response value among the multiple confidence vectors or the value after mathematical transformation of the maximum value as the predicted confidence value of the target detection network.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the position classification method of the target detection network according to any one of claims 1 to 4 are implemented.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the location classification method of the target detection network according to any one of claims 1 to 4 are implemented.
10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the location classification method of the target detection network according to any one of claims 1 to 4 are implemented.