Adaptive feature enhancement method, device, equipment and readable storage medium

Abstract

The application provides a self-adaptive feature enhancement method, device and equipment and a readable storage medium. The method comprises the following steps: performing multi-dimensional feature extraction on input image data to obtain a multi-dimensional feature map, performing fusion enhancement on the multi-dimensional feature map to generate a multi-scale feature map; inputting the multi-scale feature map into a fountain feature enhancement module, performing dimension reduction convolution module calculation at each scale to obtain the class prediction confidence of each detection unit; determining an overall loss function based on the class prediction confidence of each detection unit, and determining the detection result of the input image data according to the overall loss function. The application can effectively improve the problem of insufficient high-dimensional semantic feature point feature cohesion and make the sub-optimal detection unit pay more attention to the target region features in the receptive field and enhance the feature expression, thereby improving the image target detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly relates to a self-adaptive feature enhancement method and device, equipment and a readable storage medium. BACKGROUND

[0002] The existing target detection method based on deep convolutional neural network has achieved great success under the support of big data and rapidly developing deep learning technology, so that the precision of model detection target is greatly improved, but there are still many problems in practical application, which are limited by many aspects, such as the shape and size difference of similar targets or different targets, sample distribution, complex background and the like. The current deep convolutional neural network target detection model has many deficiencies, among which, an inherent defect of an existing detector has not been solved, that is, the feature extraction structure used by the existing deep convolutional network model is too modular, so that the network cannot intelligently extract features of the target in the detection unit. SUMMARY

[0003] The present application provides a self-adaptive feature enhancement method, device, equipment and readable storage medium, which solves the technical problem that the existing target detection method cannot extract features of the target in the detection unit, so that the detection precision is not high.

[0004] The present application provides a self-adaptive feature enhancement method, comprising:

[0005] Multi-dimensional feature extraction is performed on the input image data to obtain a multi-dimensional feature map, and the multi-dimensional feature map is fused and enhanced to generate a multi-scale feature map;

[0006] The multi-scale feature map is input into a fountain feature enhancement module, and the class prediction confidence of each detection unit is obtained through the calculation of a dimension reduction convolution module at each scale;

[0007] The overall loss function is determined based on the class prediction confidence of each detection unit, and the detection result of the input image data is determined according to the overall loss function.

[0008] According to the self-adaptive feature enhancement method provided by the present application, the multi-scale feature map is input into the fountain feature enhancement module, and the class prediction confidence of each detection unit is obtained through the calculation of a dimension reduction convolution module at each scale.

[0009] The multi-scale feature map is input into a fountain feature enhancement module to obtain a classification score;

[0010] According to the fountain map generation algorithm, the classification score and a preset detection point anchor frame type, a fountain map is obtained.

[0011] Based on the fountain fusion module and the fountain map, the multi-scale feature map is subjected to feature deconstruction and reconstruction to obtain a detection layer feature map, and a value of the detection layer feature map on a two-dimensional plane corresponds to a class prediction confidence of each detection unit.

[0012] According to the adaptive feature enhancement method provided by the application, the feature deconstruction and reconstruction of the detection unit vector based on different kernel template shapes to obtain the detection layer feature map comprises:

[0013] The non-background detection unit in the fountain map is operated, and a target value of the non-background detection unit is a label of a predicted class of the non-background detection unit.

[0014] The feature deconstruction and reconstruction of the detection unit vector based on different kernel template shapes to obtain the detection layer feature map.

[0015] According to the adaptive feature enhancement method provided by the application, the determination of the overall loss function based on the class prediction confidence of each detection unit comprises:

[0016] The first loss function L is determined based on the formula L c (p t )=-(1-p t ) γ log(p t ) where L c is the first loss function, p t is the class prediction confidence of each detection unit, and γ is a hyperparameter.

[0017] The second loss function is determined according to the class identifier, the predicted true value vector and the model predicted value vector.

[0018] The overall loss function is determined based on the first loss function and the second loss function.

[0019] According to the adaptive feature enhancement method provided by the application, the input of the multi-scale feature map into the fountain feature enhancement module and the calculation of each dimension through the dimension reduction convolution module to obtain the class prediction confidence of each detection unit further comprises:

[0020] The class prediction confidence of each detection unit is determined based on the formula Γ=F sigmoid (Y cm [n](x)) where Γ is the class prediction probability of each detection unit, n is the number of stacked convolution modules, Y cm [n] is an n-convolution module function nest, and sigmoid is a gate function.

[0021] According to the adaptive feature enhancement method provided by the application, the multi-dimensional feature map is fused and enhanced to generate a multi-scale feature map, which comprises the following steps:

[0022] The multi-dimensional feature map is fused and enhanced to generate a multi-scale feature map.

[0023] The application further provides an adaptive feature enhancement device, which comprises the following modules:

[0024] A multi-scale feature map generation module is configured to perform multi-dimensional feature extraction on input image data to obtain a multi-dimensional feature map, and fuse and enhance the multi-dimensional feature map to generate a multi-scale feature map.

[0025] A class prediction confidence determination module is configured to input the multi-scale feature map into a fountain feature enhancement module, and obtain the class prediction confidence of each detection unit through the calculation of a dimension reduction convolution module at each scale.

[0026] A detection result determination module is configured to determine a global loss function based on the class prediction confidence of each detection unit, and determine the detection result of the input image data according to the global loss function.

[0027] The application further provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the adaptive feature enhancement method according to any one of the above-mentioned methods when executing the program.

[0028] The application further provides a non-transitory computer readable storage medium, which stores a computer program, wherein the computer program is executable on a processor to implement the adaptive feature enhancement method according to any one of the above-mentioned methods.

[0029] The application further provides a computer program product, which comprises a computer program, wherein the computer program is executable on a processor to implement the adaptive feature enhancement method according to any one of the above-mentioned methods.

[0030] The adaptive feature enhancement method, device, equipment and readable storage medium provided by the application, through the multi-dimensional feature extraction of the feature extraction module on the input image data, the multi-dimensional feature map is obtained, and then the multi-dimensional feature map is fused and enhanced through the feature fusion enhancement module, and the multi-scale feature map is generated, the multi-scale feature map is taken as the input of the fountain feature enhancement module proposed by the application, in each scale, the calculation of a dimension reduction convolution module is carried out, and the output is the confidence prediction value of each detection unit in each category, and the maximum value is taken to obtain the classification map, and finally the detection result of the input image data is determined according to the overall loss function, aiming at the problem of excessive structuralization of the feature extraction structure of the existing deep convolutional neural network target detector, the problem of insufficient feature cohesion of high-dimensional semantic feature points can be effectively improved, the suboptimal detection unit pays more attention to the target area features in the receptive field, and the feature expression is enhanced, and the image target detection precision is improved. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0032] Figure 1 is one of the flowcharts of the adaptive feature enhancement method provided by the application;

[0033] Figure 2 is the second flowchart of the adaptive feature enhancement method provided by the application;

[0034] Figure 3 is the structural schematic diagram of the adaptive feature enhancement device provided by the application;

[0035] Figure 4 is the structural schematic diagram of the electronic equipment provided by the application. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme in the application will be described clearly and completely in combination with the drawings in the application. Obviously, the described embodiments are part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0037] The adaptive feature enhancement method of the application will be described below in combination with Figures 1-2

[0038] Please refer to​Figure 1 The application provides an adaptive feature enhancement method, comprising:

[0039] In step 100, multi-dimensional feature extraction is performed on input image data to obtain a multi-dimensional feature map, and the multi-dimensional feature map is fused and enhanced to generate a multi-scale feature map.

[0040] Specifically, the existing target detector comprises a feature extraction module, a feature fusion and enhancement module, a detection module and a post-processing module. Compared with the existing target detector, the target detector in the adaptive feature enhancement method provided in the embodiment comprises a feature backbone extraction network, a pyramid feature fusion part, a fountain feature enhancement module and a rotation detection module. The fountain feature enhancement module is the main difference between the adaptive feature enhancement method provided in the embodiment and the prior art. The feature backbone extraction network is used to perform multi-dimensional feature extraction on input image data to obtain a multi-dimensional feature map, and then the pyramid feature fusion part is used to fuse and enhance the multi-dimensional feature map to generate a multi-scale feature map.

[0041] In step 200, the multi-scale feature map is input into the fountain feature enhancement module, and the class prediction confidence of each detection unit is obtained through the calculation of a dimension reduction convolution module at each scale.

[0042] Specifically, the multi-scale feature map is taken as the original input, and at each scale, the calculation of a dimension reduction convolution module is performed to output the class prediction probability of each detection unit, and the dimension is [cls, w, h]. cls is the number of classes predicted by the model. In this way, the confidence prediction value of each detection unit in each class is obtained, and the maximum value is taken to obtain a CS (classification map). The dimension of the classification map is [1, w, h], and each value in the two-dimensional plane is the class prediction confidence of the corresponding detection unit.

[0043] In step 300, the overall loss function is determined based on the class prediction confidence of each detection unit, and the detection result of the input image data is determined according to the overall loss function.

[0044] Specifically, the first loss function L c of the fountain feature enhancement module can be determined by the class prediction confidence of each detection unit and a hyperparameter; the second loss function L r of the fountain feature enhancement module can be determined by the class label, the predicted true value vector and the model prediction value vector; and the rotation detection module loss can be determined by the number of positive samples of the class prediction branch, the number of positive samples in the position information regression module, the value after the anchor point represented by the preset is judged as a positive sample or a negative sample (the positive sample is 1 and the negative sample is 0), the class prediction confidence and a balance adjustment coefficient, L rThe calculation manner of L c is the same as that of the first part, and the overall loss function Loss is L c +L r , and finally the detection result of the input image data is determined according to the overall loss function.

[0045] The embodiment extracts multi-dimensional features from the input image data through the feature extraction module to obtain a multi-dimensional feature map, and then performs fusion enhancement on the multi-dimensional feature map through the feature fusion enhancement module to generate a multi-scale feature map. The multi-scale feature map is input into the fountain feature enhancement module proposed by the application. In each scale, it will be calculated through a dimension reduction convolution module, and the output is the confidence prediction value of each detection unit in each category. After taking the maximum value, a classification map can be obtained. Finally, the detection result of the input image data is determined according to the overall loss function. In view of the problem of excessive structuralization of the feature extraction structure of the existing deep convolutional neural network target detector, the problem of insufficient feature cohesion of high-dimensional semantic feature points can be effectively improved, so that the suboptimal detection unit pays more attention to the target region features in the receptive field and enhances the feature expression, thereby improving the image target detection accuracy.

[0046] Please refer to Figure 2 In one embodiment, the adaptive feature enhancement method provided by the embodiment of the application can further include:

[0047] Step 210: inputting the multi-scale feature map into the fountain feature enhancement module to obtain a classification score;

[0048] Step 220: obtaining a fountain map according to the fountain map generation algorithm, the classification score, and a preset detection point anchor frame type.

[0049] Step 230: performing feature deconstruction and reconstruction on the multi-scale feature map based on the fountain fusion module and the fountain map to obtain a detection layer feature map. The value of the detection layer feature map on a two-dimensional plane corresponds to the class prediction confidence of each detection unit.

[0050] The above step 230 includes:

[0051] Step 231: performing operation on a non-background detection unit in the fountain map. The target value of the non-background detection unit is the label of the predicted category of the non-background detection unit.

[0052] Step 232: performing feature deconstruction and reconstruction on the vector of the detection unit based on different kernel template shapes to obtain a detection layer feature map.

[0053] Specifically, the multi-scale feature map is used as the input of the fountain feature enhancement module, which has a shape of [Scale, Batch, Channel, w, h], representing the feature pyramid level label, batch size (the sample size used in one iteration), the number of feature map channels, the width of the feature map, and the height of the feature map, respectively. Each feature point in the multi-scale feature map is an independent detection unit. The core technology of the fountain feature enhancement module is the FMG (Fountain Map Generation) algorithm and the FFM (Fountain Fusion Module).

[0054] The multi-scale feature map is used as the original input, and at each scale, a dimension reduction convolution module is used to calculate the class prediction probability of each detection unit, which has a dimension of [cls, w, h]. cls is the number of classes predicted by the model. In this way, we get the confidence prediction value of each detection unit in each class. After taking the maximum value, we get the CS (classification map), which has a dimension of [1, w, h]. Each value on the two-dimensional plane is the class prediction confidence of the corresponding detection unit.

[0055] After the FMG algorithm, the fountain map of each type of anchor point is obtained. The input of the FFM is the feature map at each scale, which has a dimension of [H, W, F], representing the width and height of the detection layer feature map, and the number of feature channels. Then, based on the generated fountain map of each type of anchor point, the feature is deconstructed and reconstructed. The FFM only operates on the non-background detection units in the fountain map of each type of anchor point. The non-background units in the fountain map of each type of anchor point include not only the detection units of the target of interest, but also the detection units with high confidence. Because the FFM is designed to enhance the features of suboptimal detection units, the value of the non-background units in the fountain map of each type of anchor point is the label of the predicted class. The linear wrapper in the FFM can use different kernel templates to reconstruct the feature vector of the detection unit. First, an encoder is used to expand the dimension, and then a decoder is used to reduce the dimension to Fx1 and input it into the new feature map. The feature vector of the adjacent detection unit is directly transmitted to the new feature map through the jump link.

[0056] The output of the FFM is then stacked with the original input in the channel dimension, and the number of channels is reduced to the same as the input vector through 1x1 convolution calculation to obtain the detection layer feature map. The value of the detection layer feature map on the two-dimensional plane corresponds to the class prediction confidence of each detection unit.

[0057] This embodiment uses the fountain fusion module and the fountain map to deconstruct and reconstruct the multi-scale feature map to obtain the detection layer feature map. It effectively improves the problem of insufficient feature cohesion of high-dimensional semantic feature points by addressing the over-structured feature extraction structure of existing deep convolutional neural network target detectors. This allows the suboptimal detection unit to focus more on the target region features within the receptive field and enhance their feature representation, thereby improving the accuracy of image target detection.

[0058] In one embodiment, the adaptive feature enhancement method provided in this application may further include:

[0059] Step 310, based on formula L c (p t )=-(1-p t ) γ log(p t Determine the first loss function, L c Let p be the first loss function. t The category prediction confidence level for each of the detection units is given, where γ is a hyperparameter.

[0060] Step 320: Determine the second loss function based on the category identifier, the predicted true value vector, and the model predicted value vector;

[0061] Step 330: Determine the overall loss function based on the first loss function and the second loss function.

[0062] Specifically, the first loss function L of the fountain feature enhancement module c It can be calculated using Formula 1, where L c For the first loss function, p t The category prediction confidence of each detection unit, γ is a hyperparameter; the second loss function L of the fountain feature enhancement module. r The values ​​can be calculated using formulas 2 and 3, where u is the class identifier, v is the predicted ground truth vector, and t is the model predicted value vector. The rotation detection module loss can be calculated using formula 4, where N... cls N represents the number of positive samples in the category prediction branch. reg p represents the number of positive samples in the location information regression module. i * This represents the value after the pre-set anchor point is determined as a positive or negative sample; a positive sample is 1, and a negative sample is 0. i Let L be the confidence level for class prediction, λ be the balance adjustment coefficient, and L be the confidence level for class prediction. r The calculation method is the same as L in the first part. c Similarly, the overall loss function Loss is L. c +L r .

[0063] L c (p t )=-(1-p t ) γ log(p t ) Formula 1

[0064]

[0065]

[0066]

[0067] This embodiment calculates the loss function of the fountain feature enhancement module and adjusts the fountain feature enhancement module based on the detection results of the fountain feature enhancement module to further enhance the feature representation of the target area and improve the accuracy of image target detection.

[0068] In one embodiment, the adaptive feature enhancement method provided in this application may further include:

[0069] Step 201, based on the formula Γ=F sigmoid (Y cm [n](x)) determines the class prediction confidence of each detection unit, Γ is the class prediction probability of each detection unit, n is the number of stacked convolutional modules, and Y cm [n] represents n nested convolutional module functions, and sigmoid is the gate function.

[0070] Specifically, the multi-scale feature map, as the raw output of the fountain feature enhancement module, undergoes calculation by a dimensionality reduction convolutional module at each scale. The output is the class prediction probability for each detection unit, denoted as Γ, which can be calculated using Equation 5, where Γ is the class prediction probability of each detection unit, n is the number of stacked convolutional modules, and Y... cm [n] represents n nested convolutional module functions, and sigmoid is the gate function. The dimension of Γ is [cls, w, h], where cls is the number of categories predicted by the model. In this way, we obtain the confidence prediction value of each detection unit in each category. Taking the maximum value, we can obtain the CS (classification map). The dimension of the classification map is [1, w, h], and each value on the two-dimensional plane is the category prediction confidence of the corresponding detection unit.

[0071] Γ=F sigmoid (Y cm Formula 5

[0072] The embodiment calculates the category prediction confidence of each detection unit, and can effectively improve the problem of insufficient high-dimensional semantic feature point feature cohesion, so that the suboptimal detection unit pays more attention to the target region features in the receptive field and enhances the feature expression, thereby improving the image target detection precision.

[0073] In one embodiment, the adaptive feature enhancement method provided by the embodiment of the application further includes:

[0074] In step 110, the multi-dimensional feature map is fused and enhanced to generate a multi-scale feature map.

[0075] Specifically, the target detector in the adaptive feature enhancement method provided by the embodiment includes four parts: 1, a feature backbone extraction network (also referred to as a model skeleton network), which is used to receive image data as the model input, uses a stacked convolutional neural network group to perform feature extraction on the multi-dimensional data, and the feature extraction result is usually multi-level, so as to ensure the richness of the extracted semantic features and cope with the multi-scale changes of the target; 2, a pyramid feature fusion part (also referred to as a model neck), which can fuse and enhance the multi-dimensional feature map extracted in the feature extraction module, fuse the rich semantic features of the high-dimensional feature layer and the rich spatial information of the low-dimensional feature layer together, and generate a feature map with stronger representation capability, i.e., the multi-scale feature map in the embodiment; 3, a FFEM (Fountain Feature Enhancement Module), which is the main difference point of the adaptive feature enhancement method provided by the embodiment compared with the prior art; and 4, a rotation detection module and a detection result output module.

[0076] The multi-dimensional feature map is fused and enhanced by the pyramid feature fusion part to obtain the original input multi-scale feature map of the FFEM, thereby providing a data basis for the subsequent feature deconstruction and reconstruction.

[0077] The adaptive feature enhancement device provided by the embodiment is described below, and the adaptive feature enhancement device described below can be correspondingly referred to the adaptive feature enhancement method described above.

[0078] Please refer to Figure 3 The application further provides an adaptive feature enhancement device, which includes:

[0079] The multi-scale feature map generation module 301 is configured to perform multi-dimensional feature extraction on the input image data to obtain a multi-dimensional feature map, fuse and enhance the multi-dimensional feature map, and generate a multi-scale feature map.

[0080] The category prediction confidence determination module 302 is configured to input the multi-scale feature map into a fountain feature enhancement module, and obtain category prediction confidence of each detection unit through dimension reduction convolution module calculation at each scale.

[0081] The detection result determination module 303 is configured to determine an overall loss function based on the category prediction confidence of each detection unit, and determine a detection result of the input image data according to the overall loss function.

[0082] Optionally, the category prediction confidence determination module comprises:

[0083] The classification score determination unit is configured to input the multi-scale feature map into the fountain feature enhancement module to obtain a classification score.

[0084] The fountain map determination unit is configured to obtain a fountain map according to a fountain map generation algorithm, the classification score and a preset detection point anchor frame type.

[0085] The detection layer feature map determination unit is configured to perform feature deconstruction and reconstruction on the multi-scale feature map based on a fountain fusion module and the fountain map to obtain a detection layer feature map, wherein a value of the detection layer feature map on a two-dimensional plane corresponds to the category prediction confidence of each detection unit.

[0086] Optionally, the detection layer feature map determination unit comprises:

[0087] The non-background detection unit operation unit is configured to perform operation on a non-background detection unit in the fountain map, wherein a target value of the non-background detection unit is a label of a predicted category of the non-background detection unit.

[0088] The detection layer feature map determination unit is configured to perform feature deconstruction and reconstruction on a vector of a detection unit based on different kernel template shapes to obtain a detection layer feature map.

[0089] Optionally, the detection result determination module comprises:

[0090] The first loss function determination unit is configured to determine a first loss function L c based on a formula L t =-(1-p t ) γ log(p t ) , wherein L c is the first loss function, p t is the category prediction confidence of each detection unit, and γ is a hyperparameter.

[0091] The second loss function determination unit is configured to determine a second loss function according to a category identifier, a predicted true value vector and a model predicted value vector.

[0092] The overall loss function determination unit is used to determine the overall loss function based on the first loss function and the second loss function.

[0093] Optionally, the category prediction confidence determination module further includes:

[0094] Category prediction confidence determination unit, used based on formula Γ = F sigmoid (Y cm [n](x)) determines the class prediction confidence of each detection unit, Γ is the class prediction probability of each detection unit, n is the number of stacked convolutional modules, and Y cm [n] represents n nested convolutional module functions, and sigmoid is the gate function.

[0095] Optionally, the multi-scale feature map generation module includes:

[0096] The fusion enhancement unit is used to fuse and enhance the multi-dimensional feature map to generate a multi-scale feature map.

[0097] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute an adaptive feature enhancement method.

[0098] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program being stored in a non-transitory computer readable storage medium, and the computer program, when executed by a processor, enables a computer to perform the adaptive feature enhancement method provided by the above-mentioned methods.

[0100] In yet another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program, when executed by a processor, enables a computer to perform the adaptive feature enhancement method provided by the above-mentioned methods.

[0101] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0102] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a number 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.

[0103] 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 they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An adaptive feature enhancement method, characterized in that, include: Multidimensional feature maps are obtained by extracting multidimensional features from the input image data, and the multidimensional feature maps are then fused and enhanced to generate multi-scale feature maps. The multi-scale feature map is input into the fountain feature enhancement module, and after calculation by the dimension reduction convolution module at each scale, the category prediction confidence of each detection unit is obtained. The overall loss function is determined based on the category prediction confidence of each detection unit, and the detection result of the input image data is determined based on the overall loss function. The multi-scale feature map is input into the fountain feature enhancement module, and after calculation by the dimensionality reduction convolution module at each scale, the category prediction confidence of each detection unit is obtained, including: The multi-scale feature map is input into the fountain feature enhancement module to obtain the classification score; The fountain map is obtained based on the fountain map generation algorithm, the classification score, and the preset detection point anchor box type; Based on the fountain fusion module and the fountain map, the multi-scale feature map is deconstructed and reconstructed to obtain the detection layer feature map. The value of the detection layer feature map on the two-dimensional plane corresponds to the category prediction confidence of each detection unit. The feature map obtained by deconstructing and reconstructing the multi-scale feature map based on the fountain fusion module and the fountain map to obtain the detection layer feature map includes: The operation is performed on the non-background detection units in the fountain map, where the target value of the non-background detection unit is the label of the predicted category of the non-background detection unit; Based on different kernel template shapes, the vectors of the detection units are deconstructed and reconstructed to obtain the feature map of the detection layer.

2. The adaptive feature enhancement method according to claim 1, characterized in that, The determination of the overall loss function based on the category prediction confidence of each of the detection units includes: Based on formula Determine the first loss function. Let the first loss function be... Calculate the confidence level for the category of each of the detection units. For hyperparameters; The second loss function is determined based on the category identifier, the predicted true value vector, and the model predicted value vector. The overall loss function is determined based on the first loss function and the second loss function.

3. The adaptive feature enhancement method according to claim 1, characterized in that, The step of inputting the multi-scale feature map into the fountain feature enhancement module, and calculating the category prediction confidence of each detection unit through the dimensionality reduction convolution module at each scale, further includes: Based on formula Determine the category prediction confidence level for each detection unit. The probability of class prediction for each of the detection units is given, where n is the number of stacked convolutional modules. There are n nested convolutional module functions, with sigmoid as the gate function. .

4. The adaptive feature enhancement method according to claim 1, characterized in that, The step of fusing and enhancing the multi-dimensional feature maps to generate multi-scale feature maps includes: The multidimensional feature maps are fused and enhanced to generate multi-scale feature maps.

5. An adaptive feature enhancement device, characterized in that, include: The multi-scale feature map generation module is used to extract multi-dimensional features from the input image data to obtain a multi-dimensional feature map, and to fuse and enhance the multi-dimensional feature map to generate a multi-scale feature map. The category prediction confidence determination module is used to input the multi-scale feature map into the fountain feature enhancement module, and obtain the category prediction confidence of each detection unit by the calculation of the dimension reduction convolution module at each scale. The detection result determination module is used to determine the overall loss function based on the category prediction confidence of each detection unit, and to determine the detection result of the input image data according to the overall loss function; The multi-scale feature map is input into the fountain feature enhancement module, and after calculation by the dimensionality reduction convolution module at each scale, the category prediction confidence of each detection unit is obtained, including: The multi-scale feature map is input into the fountain feature enhancement module to obtain the classification score; The fountain map is obtained based on the fountain map generation algorithm, the classification score, and the preset detection point anchor box type; Based on the fountain fusion module and the fountain map, the multi-scale feature map is deconstructed and reconstructed to obtain the detection layer feature map. The value of the detection layer feature map on the two-dimensional plane corresponds to the category prediction confidence of each detection unit. The feature map obtained by deconstructing and reconstructing the multi-scale feature map based on the fountain fusion module and the fountain map to obtain the detection layer feature map includes: The operation is performed on the non-background detection units in the fountain map, where the target value of the non-background detection unit is the label of the predicted category of the non-background detection unit; Based on different kernel template shapes, the vectors of the detection units are deconstructed and reconstructed to obtain the feature map of the detection layer.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the adaptive feature enhancement method as described in any one of claims 1 to 4.

7. 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, it implements the adaptive feature enhancement method as described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the adaptive feature enhancement method as described in any one of claims 1 to 4.

Images