A target detection method and system based on adaptive feature extraction and multi-scale enhancement

Abstract

The application relates to a target detection method and system based on adaptive feature extraction and multi-scale enhancement. The method comprises the following steps: performing standardization preprocessing on an original image to obtain a to-be-detected original image; inputting the to-be-detected original image into a dilated convolution block with a parallel structure to capture feature information of the to-be-detected original image at different scales; dynamically adjusting the contribution degree of each channel feature through a target normalization method to obtain an optimal feature vector; scaling the optimal feature vector to obtain compressed target attention weights; using a channel attention gate method to calculate a gating signal for the target attention weights and generate a fused feature map; extracting deep features through a multi-scale enhancement feature pyramid network; and outputting a target detection result through a classifier and a regressor. The method can improve the accuracy and effectiveness of target detection on targets at different scales and enhance the detection capability of targets in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and in particular to a target detection method and system based on adaptive feature extraction and multi-scale enhancement. Background Technology

[0002] Object detection is a core task in computer vision, aiming to identify and locate objects of interest in images or videos. Object detection not only needs to identify the category of objects in an image but also accurately label their locations. It has applications in numerous fields such as intelligent security, intelligent robotics, and intelligent transportation. Traditional object detection typically uses a convolutional neural network (CNN) to extract features from the input image, obtaining key information. Then, bounding boxes are used to label the location of target objects; a bounding box is a rectangle that accurately delineates the scope of the target object. Finally, the category of the target object is determined by the output of the convolutional neural network.

[0003] However, in the field of object detection, traditional techniques have many limitations when dealing with targets of different scales. These include: difficulty in effectively adapting to changes in target scale during feature extraction, resulting in insufficient detection capability for targets at extreme scales; inefficiency in feature fusion methods at different scales, failing to fully integrate the advantages of features at different levels; and a lack of flexibility in loss functions in balancing the difficulty of detecting targets at different scales, affecting detection accuracy and robustness.

[0004] Therefore, traditional target detection methods often have low accuracy and effectiveness on targets of different scales, and they also have insufficient target detection capabilities in complex scenarios. Summary of the Invention

[0005] Therefore, in order to solve the above-mentioned technical problems, a target detection method and system based on adaptive feature extraction and multi-scale enhancement is provided, which can improve the accuracy and effectiveness of target detection on targets at different scales and enhance the target detection capability in complex scenes.

[0006] A target detection method based on adaptive feature extraction and multi-scale enhancement, the method comprising:

[0007] The original input image is obtained, and the original image is subjected to standardization preprocessing to obtain the original image to be detected;

[0008] The original image to be detected is input into a parallel structured dilated convolutional block. The convolutional layers with different dilation rates in the dilated convolutional block capture the feature information of the original image to be detected at different scales, extract the image features, and obtain feature maps for each channel.

[0009] Based on the image features, a target normalization method is determined. The contribution of each channel feature is dynamically adjusted through the target normalization method to obtain the optimal feature vector. The optimal feature vector is then scaled to obtain the compressed target attention weight. A channel attention gating method is used to calculate the gating signal for the target attention weight, and a fused feature map is generated based on the gating signal and the channel feature map.

[0010] The fused feature map is input into a multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the middle layer, thereby achieving multi-scale feature fusion and extracting deep features.

[0011] The deep features are input into the classifier and regressor, and the cross-union ratio threshold is dynamically adjusted using the scale-adaptive cross-union ratio loss function to output the target detection results.

[0012] In one embodiment, the input raw image is acquired, and the raw image is subjected to normalization preprocessing to obtain the raw image to be detected, including:

[0013] Obtain the input original image and obtain the pixel matrix of the original image;

[0014] The pixel matrix is ​​normalized according to predetermined data standardization rules to obtain the original image to be detected.

[0015] In one embodiment, the original image to be detected is input into a dilated convolutional block of a parallel structure. Convolutional layers with different dilation rates within the dilated convolutional block capture feature information of the original image at different scales, extracting image features to obtain feature maps for each channel, including:

[0016] The original image to be detected is input as a tensor into a dilated convolution block of a parallel structure;

[0017] Parallel feature extraction is performed using convolutional layers with different dilation rates in the dilated convolutional block to obtain feature maps at various scales and receptive fields.

[0018] The feature maps are fused along the channel dimension using a feature concatenation method to obtain fused channel feature maps.

[0019] In one embodiment, a target normalization method is determined based on the image features. This method dynamically adjusts the contribution of each channel's features to obtain an optimal feature vector, including:

[0020] Each of the channel feature maps is normalized in spatial dimension using various normalization methods to obtain the normalization result;

[0021] The optimal normalization method is selected as the target normalization method based on the normalization result using the activation function.

[0022] The contribution of each channel feature is dynamically adjusted using the target normalization method to obtain the optimal feature vector.

[0023] In one embodiment, scaling the optimal feature vector to obtain compressed target attention weights includes:

[0024] The optimal feature vector is scaled using a set of learnable weight parameters and then scaled again using a parameterless normalization technique to obtain a compressed feature vector.

[0025] A specific scalar is introduced into the compressed feature vector to normalize the scale, and scale normalization is performed to obtain the compressed target attention weights.

[0026] In one embodiment, a channel attention gating method is used to calculate a gating signal for the target attention weights, and a fused feature map is generated based on the gating signal and the channel feature map, including:

[0027] The target attention weights are applied to a channel attention gating method using an activation function to calculate the gate control signal.

[0028] The channel feature map is weighted using the target attention weights to obtain a weighted feature map;

[0029] The gated signal is multiplied element by element by the weighted feature map to obtain the result of the multiplication;

[0030] The result of the multiplication is summed along the first dimension to obtain the fused feature map.

[0031] In one embodiment, the fused feature map is input into a multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the intermediate layer, achieving multi-scale feature fusion and extracting deep features, including:

[0032] The fused feature map is input into a multi-scale augmented feature pyramid network, and the features in the fused feature map are scaled by a linear scaling layer in the multi-scale augmented feature pyramid network.

[0033] The scaled features are then added pixel-by-pixel to the original features and input into the fusion layer of the multi-scale enhanced feature pyramid network.

[0034] The fusion layer performs convolution operations on the input features to extract and achieve multi-scale feature fusion, thereby obtaining the deep features of the fused feature map.

[0035] In one embodiment, the method further includes:

[0036] Determine the ground truth bounding box and the predicted bounding box of the target in the original image to be detected, and calculate the intersection-union ratio of the ground truth bounding box and the predicted bounding box;

[0037] Calculate the Euclidean distance between the center points of the ground truth bounding box and the predicted bounding box, and calculate the diagonal length of the minimum closure region covering the ground truth bounding box and the predicted bounding box;

[0038] The target size function is defined based on the diagonal length, and the scale-adaptive cross-union ratio loss function is obtained according to the cross-union ratio, Euclidean distance, diagonal length, and target size function.

[0039] In one embodiment, the deep features are input into a classifier and a regressor, and the cross-union ratio (CUI) threshold is dynamically adjusted using a scale-adaptive CUI loss function to output target detection results, including:

[0040] The deep features are input into the classifier and regressor. The classifier predicts the target category based on the deep features to obtain the probability distribution of the target belonging to different categories.

[0041] The target position and size are predicted using the regressor, and the prediction results are obtained.

[0042] The probability distribution, location, and size prediction results are output as target detection results.

[0043] A target detection system based on adaptive feature extraction and multi-scale enhancement, the system comprising:

[0044] The image processing module is used to acquire the input raw image and perform standardization preprocessing on the raw image to obtain the raw image to be detected;

[0045] An adaptive feature extraction module is used to input the original image to be detected into a parallel structured dilated convolution block, and capture the feature information of the original image to be detected at different scales through convolutional layers with different dilation rates in the dilated convolution block, extract image features, and obtain feature maps of each channel.

[0046] The feature processing and fusion module is used to determine the target normalization method based on the image features, dynamically adjust the contribution of each channel feature through the target normalization method to obtain the optimal feature vector, scale the optimal feature vector to obtain the compressed target attention weight, calculate the gating signal for the target attention weight using the channel attention gating method, and generate the fused feature map based on the gating signal and the channel feature map.

[0047] The multi-scale feature enhancement module is used to input the fused feature map into the multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the middle layer, thereby achieving multi-scale feature fusion and extracting deep features.

[0048] The target detection module is used to input the deep features into the classifier and regressor, dynamically adjust the cross-union ratio threshold using the scale-adaptive cross-union ratio loss function, and output the target detection results.

[0049] The aforementioned target detection method and system based on adaptive feature extraction and multi-scale enhancement uses parallel dilated convolution sequences to extract features from images through progressively increasing dilation rates, effectively expanding the network's receptive field and enabling better capture of contextual information in the image, thus improving the accuracy of large target detection. The adaptive normalization method extracts channel attention weights, dynamically adjusting the importance of feature channels and enhancing sensitivity to key features, contributing to improved overall detection performance. The use of a multi-scale enhanced feature pyramid network fully integrates feature maps of different scales, effectively detecting targets of varying sizes. Furthermore, the use of a scale-adaptive intersection-union loss function reasonably balances the loss contribution of targets of different sizes, improving the overall performance and adaptability of the target detection method. Attached Figure Description

[0050] Figure 1 This is an application environment diagram of a target detection method based on adaptive feature extraction and multi-scale enhancement in one embodiment.

[0051] Figure 2 This is a flowchart illustrating a target detection method based on adaptive feature extraction and multi-scale enhancement in one embodiment;

[0052] Figure 3 This is a schematic diagram of a multi-scale enhanced feature pyramid network structure and a bidirectional fusion module in one embodiment.

[0053] Figure 4 This is a schematic diagram of the scale-adaptive intersection-union ratio loss function in one embodiment;

[0054] Figure 5This is a block diagram of a target detection system based on adaptive feature extraction and multi-scale enhancement in one embodiment.

[0055] Figure 6 This is a schematic diagram of the adaptive feature extraction module in one embodiment;

[0056] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] The target detection method based on adaptive feature extraction and multi-scale enhancement provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes computer device 110. Computer device 110 can acquire the input raw image and perform standardized preprocessing on the raw image to obtain the raw image to be detected. Computer device 110 can input the raw image to be detected into a parallel structured dilated convolutional block, capturing feature information of the raw image at different scales through convolutional layers with different dilation rates, extracting image features, and obtaining feature maps for each channel. Computer device 110 can determine a target normalization method based on the image features, dynamically adjusting the contribution of each channel feature through the target normalization method to obtain the optimal feature vector; and scale the optimal feature vector to obtain... The compressed target attention weights are used to calculate a gating signal using a channel attention gating method. A fused feature map is then generated based on the gating signal and the channel feature map. The computer device 110 can input the fused feature map into a multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the intermediate layer, achieving multi-scale feature fusion and extracting deep features. The computer device 110 can then input these deep features into a classifier and a regressor, dynamically adjusting the cross-union ratio (CUI) threshold using a scale-adaptive cross-union loss function, and outputting the target detection result. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, and other devices.

[0059] In one embodiment, such as Figure 2 As shown, a target detection method based on adaptive feature extraction and multi-scale enhancement is provided, including the following steps:

[0060] Step 202: Obtain the input original image and perform standardization preprocessing on the original image to obtain the original image to be detected.

[0061] Computer equipment can acquire the input raw image and then preprocess the raw image. Specifically, in one embodiment, a target detection method based on adaptive feature extraction and multi-scale enhancement may further include an image preprocessing process, which includes: acquiring the input raw image and acquiring the pixel matrix of the raw image; and normalizing the pixel matrix according to a predetermined data standardization rule to obtain the raw image to be detected.

[0062] In this embodiment, the preprocessing method for the input raw image data may be to standardize the raw image data to ensure that the data input into the adaptive feature extraction has a consistent format and scale. The standardization operation for the raw image is based on general data preprocessing methods, including but not limited to pixel value normalization and size adjustment.

[0063] For example, I represents the original image, C represents the number of color channels, and H and W represent the height and width of the image, respectively. For ease of calculation and presentation, some data for a specific color channel in I can be shown in the following table:

[0064]

[0065]

[0066] In this embodiment, the computer device can read the original image data, obtain the image's pixel matrix and related data, such as image width W and height H, number of channels C, etc.; and normalize the pixel matrix according to a predetermined data standardization rule. For example, if the pixel value range of image I is [min val ,max val ], then through formula x c =(I-min) val ) / (max val -min val The pixel values ​​are mapped to the [0,1] interval; the image size is adjusted according to the input requirements of the object detection network, specifically by cropping, scaling, etc., to unify the image to a specific width and height, such as 256×256×3.

[0067] Step 204: Input the original image to be detected into the dilated convolution block of the parallel structure. The convolutional layers with different dilation rates in the dilated convolution block capture the feature information of the original image to be detected at different scales, extract the image features, and obtain the feature maps of each channel.

[0068] In one embodiment, the target detection method based on adaptive feature extraction and multi-scale enhancement may further include an adaptive feature extraction process, which specifically includes: inputting the original image to be detected into a parallel structure of dilated convolutional blocks in the form of tensors; performing parallel feature extraction through convolutional layers with different dilation rates in the dilated convolutional blocks to obtain feature maps at different scales and receptive fields; and fusing the feature maps in the channel dimension by feature concatenation to obtain fused feature maps for each channel.

[0069] In this approach, parallel-structured dilated convolutional blocks can extract features from preprocessed image data using a sequence of progressively dilated convolutions. For example, the input image is... Where H is the height, W is the width, and C is the number of channels; the dilated convolution is k×k, with K=3 by default; the actual receptive field size covered by the convolutional kernel with dilation rate d is RF=(k-1)·d+1; the receptive field size of multiple layers is RF. n =RF n-1 +(k-1)·d; the minimum receptive field that can cover the target can be described as +(k-1)·d;

[0070] The expansion rates of each branch are D = {d0 = 1, d1 = 2, ..., d...} n =n} is a sequence that expands stepwise. Calculate RF. n Does it satisfy the minimum receptive field? The requirement is to select the minimum number of parallel branches N such that the total receptive field of all branches is maximized.

[0071] The input image data is sequentially fed into dilated convolution branches for feature extraction, and then concatenated along the channel dimension to form the feature map. Where N is the number of branches, convolutional layers with different dilation rates can simultaneously capture feature information of the image at different scales, avoiding the problem of feature information loss or weakening caused by multiple convolution operations in traditional serial structures.

[0072] In this embodiment, the adaptive feature extraction process may specifically include: color image data after input data preprocessing. Where H is the height, W is the width, and C is the number of channels, its format is adjusted to suit the tensor of the network input, such as... Image data X is sequentially input as tensors into dilated convolutional blocks in a parallel structure for adaptive feature extraction.

[0073] In a parallel dilated convolution sequence, convolution operations are performed according to the set dilation rate and kernel size. Taking a dilated convolution layer with a dilation rate of 2 as an example, for the input feature map, the convolution operation skips one pixel in space for sampling, thereby expanding the receptive field. After parallel feature extraction by N branches of dilated convolution layers, N feature maps with different scales and receptive fields are obtained. Then, these N feature maps are fused along the channel dimension using feature concatenation. Assuming that each feature map has C channels, the number of channels in the fused feature map will become N×C, and the size of the fused feature map will be... By integrating rich feature information extracted under different dilation rates, it can better represent various features in the image, especially when processing images containing targets of different scales, exhibiting stronger adaptability and feature representation capabilities. Next, the fused feature map U can be fed into an adaptive normalization method for further processing.

[0074] Specifically, in this embodiment, the input image is For ease of calculation, we use H=5, W=5, C=3 for calculation; the minimum receptive field for target coverage is f(x) = min(RF). total ) = H + W - 1 = 5 + 5 - 1 = 9; According to the formula RF total =k0+(k0-1)×(d1-1)+(d1+(k1-1)×(d2-1))+...+(k n-1 +{k n-1 -1)) is calculated to obtain the number of branches of the convolutional layer. The receptive field size of the first branch is 3 + (3-1) × (2-1) = 5; the receptive field size of the second branch is 5 + (3 + (3-1) × (2-1)) = 9; the final number of branches of the dilated convolution is N = 2 and the dilation rate is D = {d0 = 1, d1 = 2}. In this embodiment, for the convolutional layer with dilation rate d0 = 1, according to the convolution formula... Output the pixel value of the feature map, where w mn It is the convolution kernel weight, b c It is a bias term, b c This refers to the pixel value at the corresponding location in the input feature map. Assume the convolutional kernel weights are randomly initialized within the range [-0.1, 0.1], and the bias term is initialized to 0. After this convolutional layer, the output feature map size is C1×222×222, and the number of output channels is C1=64. For a convolutional layer with an expansion rate of d1=2, its calculation formula is... However, due to the dilation rate of 2, one pixel is skipped in spatial sampling, resulting in an output feature map size of C2×220×220. The calculation method is the same for convolutional layers with different dilation rates.

[0075] Next, the N branch feature maps can be fused along the channel dimension using feature concatenation. The size of the fused feature map is [size missing]. C = [1,2,...,C], where C is the number of channels. The fused feature map will be fed into the channel attention gating for further processing.

[0076] Step 206: Determine the target normalization method based on image features. Dynamically adjust the contribution of each channel feature through the target normalization method to obtain the optimal feature vector. Scale the optimal feature vector to obtain the compressed target attention weight. Calculate the gating signal for the target attention weight using the channel attention gating method. Generate a fused feature map based on the gating signal and the channel feature map.

[0077] First, the computer device can perform normalization operations based on image features, using normalization methods from the standard library to calculate feature vectors, and take the optimal value as the gating weight.

[0078] An adaptive normalization method is applied to learn the importance of channel feature maps in different branches, dynamically adjusting the contribution of each channel feature to enhance sensitivity to key features. In this embodiment, a standardized normalization library can be uniformly adopted, including but not limited to global average pooling Fi. GAP =GAP(U)=max 1≤i≤H,1≤j≤W U(i,j), Global Max Pooling L1 norm L2 norm General methods, etc.

[0079] Global average pooling: For feature maps of different channels, a global average pooling operation is performed in the spatial dimension, converting the two-dimensional feature map into a numerical value. The calculation formula is as follows: Suppose that some pixel values ​​of the first channel feature map are as follows This example only shows the calculation for the top 4 pixels of the image; the actual calculation covers the entire H×W range. The calculation yields... This operation is performed on all channels in sequence to obtain a feature vector with dimension [C,1,1].

[0080] Global max pooling: For feature maps of different channels, perform a global max pooling operation to convert the two-dimensional feature map into a single numerical value. The calculation formula is as follows: Taking the first channel as an example again, the calculation yields... After performing this operation on all channels, the corresponding global max pooling result is obtained, which ultimately forms a feature vector with dimension [C,1,1].

[0081] Calculate the L1 norm: The formula is as follows Taking the first channel as an example, the calculation process is as follows: Finally, the L1 norm value of the channel is calculated; in the same way, the L1 norm of all channels is calculated in turn, and finally a feature vector with dimension [C,1,1] is formed.

[0082] Calculate the L2 norm: The formula is as follows Where x c (i,j) represents the pixel value at position (i,j) in the c-th channel; again, taking the 1st channel as an example, calculate... The L2 norm of all channels is calculated sequentially, and finally a feature vector with dimension [C,1,1] is formed.

[0083] Specifically, in one embodiment, the provided target detection method based on adaptive feature extraction and multi-scale enhancement may further include the process of obtaining the optimal feature vector. The specific process includes: normalizing the feature maps of each channel in terms of spatial dimension using various normalization methods to obtain the normalization result; selecting the best normalization method as the target normalization method based on the normalization result using an activation function; and dynamically adjusting the contribution of each channel feature using the target normalization method to obtain the optimal feature vector.

[0084] Computer equipment can analyze the feature map of each channel. For spatial normalization, global average pooling F can be used. GAP =GAP(U)=max 1≤i≤H,1≤j≤W U(i,j), Global Max Pooling L1 norm L2 norm General methods, etc.

[0085] Adaptively selects the optimal normalization method based on the input data, that is, automatically learns and selects the most suitable normalization method for the current task and data during training, described as U. out =ω1F GAP (U)+ω2F GMP (U)+ω3F L1 (U)+ω4F L2 (U), where the weights ω of the normalized standard library i satisfy And ω i ≥0; Using the Softmax activation function, the optimal normalization method is adaptively selected. The output of Softmax is in the range [0,1], resulting in the best feature representation. eigenvectors.

[0086] Applying the Softmax activation function, with the normalized values ​​as input, and using Softmax(x) = [0.25, 0.4, 1.0, 0.5477], the output weights are [0.1816, 0.2123, 0.3826, 0.2345]. The maximum weight is taken as the optimal normalization method to obtain the feature vector.

[0087] In one embodiment, the target detection method based on adaptive feature extraction and multi-scale enhancement may further include a process of scaling the features. The specific process includes: scaling the optimal feature vector through a set of learnable weight parameters and scaling it using a parameterless normalization technique to obtain a compressed feature vector; introducing a specific scalar into the compressed feature vector to normalize the scale, performing scale normalization to obtain the compressed target attention weights.

[0088] The resulting feature vector is passed through a set of learnable weight parameters τ = [τ1,...,τ2]. i ,...,τ c Scaling is performed using the expression z. c =τ c ·s c Suppose that the learnable weight parameter τ is a vector with dimensions matching the feature vector dimensions, and its values ​​are randomly initialized, for example, τ = [0.5, 0.6, ..., 0.4]. The optimal feature vector is assumed to be s. c If the vector z is [1.2, 1.3, ..., 1.1], then the scaled vector z... c The elements are calculated as z1 = 0.5 × 1.2 = 0.6, z2 = 0.6 × 1.3 = 0.78, and the scaled complete feature vector z is obtained by calculating these elements sequentially. c Next, feature compression is performed using a parameterless normalization technique, expressed as follows: Where, ∈=10 -8 Using the scaled feature vector z obtained just now c For example, after normalization, the complete compressed feature vector is calculated sequentially. In the process of performing scale normalization, a specific scalar is introduced. For z c The scale is standardized to avoid issues when the number of channels C is large. The scale is too small, thus ensuring the stability and effectiveness of the feature scale; ultimately, the optimal compressed channel attention weights G = (g1,...,g c ).

[0089] In one embodiment, the target detection method based on adaptive feature extraction and multi-scale enhancement may further include a feature fusion process, which specifically includes: using a channel attention gating method to apply an activation function to the target attention weights to calculate a gate control signal; using the target attention weights to weight the channel feature map to obtain a weighted feature map; multiplying the gate control signal and the weighted feature map element by element to obtain the result of the multiplication; and summing the result of the multiplication along the first dimension to obtain a fused feature map.

[0090] Weights of N branches The gating signal is calculated using the activation function tanh. This gating signal is then multiplied element-wise with the stacked feature maps U, and the results are summed along the first dimension to obtain the final fused feature map, expressed by the formula: Shape

[0091] Specifically, after normalization, the gating adaptive module... The tanh activation function is used to obtain gating weight information. The output of tanh ranges from -1 to 1, which helps maintain numerical stability during training while allowing features to be flexibly adjusted in both positive and negative directions to adapt to different feature distributions. Its expression is: in, It is the feature map after channel attention weighting. It is the normalized scale vector. and ψ c These are the gate weights and biases, respectively. The scale of each channel in the feature map is dynamically adjusted by learning these parameters. Assume the gate weights... Random initialization The bias is also randomly initialized ψ c =0.1; using the normalized eigenvectors Taking the first element as an example, we can calculate that... The result G of each channel after processing by the gating adaptive module is calculated sequentially. C This forms a new feature vector with dimensions still [N,C,1,1].

[0092] Residual connectivity and fusion: This method fuses processed features with original features using residual connectivity. Taking the first channel as an example, suppose the example element of the feature vector of the first channel of the original feature map is x. c =[0.5,0.6,...,1.1], then the new feature vector elements of the first channel after fusion are calculated as follows: This process is repeated for each channel to obtain a complete new feature map after fusion. This completes the module's processing flow for the input feature map. The fused feature map can then be passed to subsequent network modules for further processing, such as participating in convolution, pooling, or classification operations in the next layer.

[0093] Step 208: Input the fused feature map into the multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the middle layer, thereby achieving multi-scale feature fusion and extracting deep features.

[0094] like Figure 3 As shown, the multi-scale enhanced feature pyramid network uses a bidirectional fusion module to fuse the features extracted from the backbone network. The bidirectional fusion module includes a linear scaling layer and a fusion layer. Furthermore, the linear scaling layer uses upsampling and downsampling to scale the features of different layers.

[0095] The weighted fused feature map is fed into a multi-scale enhanced feature pyramid network. Through a bidirectional fusion module, feature maps of different scales are combined to construct a feature pyramid that can cover targets at multiple scales. The bidirectional fusion module includes a linear scaling layer and a fusion layer. The linear scaling layer includes, but is not limited to, linear interpolation, linear filtering, and linear transformation. The fusion layer uses a 3×3 convolution kernel and extracts deep features of the image through a convolution operation with a stride of 1.

[0096] Specifically, in one embodiment, the target detection method based on adaptive feature extraction and multi-scale enhancement may further include a multi-scale feature fusion process. The specific process includes: inputting the fused feature map into a multi-scale enhanced feature pyramid network; scaling the features in the fused feature map through a linear scaling layer in the multi-scale enhanced feature pyramid network; performing pixel-by-pixel addition on the scaled features and the original features and then inputting them into the fusion layer in the multi-scale enhanced feature pyramid network; performing convolution operations on the input features through the fusion layer to extract and achieve multi-scale feature fusion, thereby obtaining the deep features of the fused feature map.

[0097] The linear scaling layer starts with the lowest resolution feature layer and uses an upsampling operation F. up =UpSampling(F high ,α), where α is the scaling factor, F high For high-level feature maps, their resolution is progressively increased to match that of adjacent feature layers. Specifically, the upsampling process begins with feature layer F2, gradually increasing its resolution to match that of the adjacent feature layer F3. The upsampling method used is assumed to be bilinear interpolation with a scaling factor α = 1.5. Taking one channel in the feature layer as an example, let the upsampled feature map be F... upThe pixel value at position (i,j) is calculated based on the weighted average of the pixels surrounding the corresponding position in the original feature map F1; for example, for F up (10,10)=w1×F3(4,4)+w2×F3(4,5)+w3×F3(5,4)+w4×F3(5,5), where w1, w2, w3, and w4 are weights determined by the bilinear interpolation algorithm, and their sum is 1.

[0098] Next, the low-level features are downsampled using a convolutional layer with a stride of 2. down =DownSampling(F low ,β), scaling factor β=0.75, F low From the low-level feature maps, we obtain feature layers F with gradually decreasing resolution. down Then the scaled features are compared with the intermediate layer features F. orig Perform pixel-by-pixel addition operation F add =F orig +F up +F down This enhances the expressive power of the features. Specifically, the downsampling operation uses a convolutional layer with a stride of 2 for downsampling the low-level feature F4. The stride is used to control the resolution reduction by half, resulting in a feature layer with gradually decreasing resolution. The goal is to make its resolution the same as the adjacent feature layer F3. For convolutional downsampling with a stride of 2, taking one channel as an example, the convolutional kernel performs sampling calculation every 2 pixels as it slides on the feature map. Assuming the convolutional kernel size is 3×3, the formula for calculating each pixel value of the output feature map is as follows: , where w mn These are the convolution kernel weights, and F4(2i+m,2j+n) is the pixel value at the corresponding position in the input feature map; for example, for the output feature map...

[0099] The above upsampled feature layer F up Perform pixel-by-pixel addition with the intermediate layer F3; similarly, perform pixel-by-pixel addition on the downsampled feature layer F. down By performing pixel-by-pixel addition on the intermediate layer F3, the fused feature layer F is obtained. fusion F fused =Conv(F add ,K 3×3 ).

[0100] The fusion layer then employs a 3×3 convolutional layer with 64 input channels and 128 output channels to further extract and fuse features, achieving deeper feature extraction. For the output feature map... in These are the convolution kernel weights corresponding to output channel C. The input feature map F fusion The pixel value at the corresponding position.

[0101] Multiple layers of bidirectional fusion involve repeated upsampling fusion from low to high resolution and downsampling fusion from high to low resolution, pixel-wise addition, and convolutional operations in the fusion layer. Each round of fusion allows for more comprehensive information exchange between feature maps of different scales, with low-resolution feature maps gaining more detailed information and high-resolution feature maps incorporating more semantic information. After multiple layers of bidirectional fusion, a complete multi-scale feature pyramid is formed. This pyramid contains feature maps at multiple scales, each incorporating information from multiple rounds of fusion from low to high resolution and from high to low resolution. This comprehensively represents the features of objects at different scales in the image, providing rich and effective feature inputs for subsequent object detection classifiers and regressors.

[0102] Step 210: Input the deep features into the classifier and regressor, dynamically adjust the cross-union ratio threshold using the scale-adaptive cross-union ratio loss function, and output the target detection results.

[0103] In one embodiment, the target detection method based on adaptive feature extraction and multi-scale enhancement may further include a process of obtaining a scale-adaptive cross-union ratio (CURRR) loss function. The specific process includes: determining the ground truth bounding box and the predicted bounding box of the target in the original image to be detected, and calculating the CURRR between the ground truth bounding box and the predicted bounding box; calculating the Euclidean distance between the center points of the ground truth bounding box and the predicted bounding box, and calculating the diagonal length of the minimum closure region covering the ground truth bounding box and the predicted bounding box; defining a target size function based on the diagonal length, and obtaining the scale-adaptive CURRR loss function based on the CURRR, the Euclidean distance, the diagonal length, and the target size function.

[0104] In the scale-adaptive intersection-union loss function, scale weights are assigned based on the target's scale. During loss calculation, the scale weight α is used to balance the loss contribution of targets of different sizes. Small targets are given larger weights to optimize the detection accuracy of small targets while maintaining the accuracy of large target detection, thus focusing more on the detection performance of targets at different scales.

[0105] A schematic diagram of the scale-adaptive cross-union ratio loss function is shown below. Figure 4 As shown, the true bounding box of the target is B. g =[x g1 ,y g1 ,x g2 ,y g2 The predicted bounding box is B. p =[x p1 ,yp1 ,x p2 ,y p2 The intersection-over-union ratio (IoU) of the predicted bounding box and the ground truth bounding box. In this embodiment, it is assumed that the coordinate information of the ground truth bounding box and the predicted bounding box in the target detection task scenario is as follows: Ground truth bounding box B g The top left corner is (20, 30), and the bottom right corner is (60, 80); prediction box B p The top-left corner is (25, 35), and the bottom-right corner is (55, 75). Therefore, calculate the true bounding box B. g And prediction box B p The coordinates of the intersection are determined by taking the maximum value of the two boxes on the horizontal and vertical axes as the top-left corner coordinates and the minimum value as the bottom-right corner coordinates. The top-left corner coordinates of the intersection are (max(20,25),max(30,35))=(25,35), and the bottom-right corner coordinates are (min(60,55),min(80,75))=(55,75). The predicted area of ​​the box is area(B). p )=(55-25)×(75-35)=1200; Actual area of ​​the frame erea(B g )=(60-20)×(80-30)=2000; the intersection area is area(B p ∩B g )=(55-25)×(75-35)=1200; Union area area(B p ∪B g =area(B) p )+area(B g )-area(B p ∩B g ) = 1200 + 2000 - 1200 = 2000; then the intersection-union ratio Next, the Euclidean distance d between the center point of the predicted bounding box and the center point of the ground truth bounding box is calculated as d = (B p B g The calculation formula is: Where (x) p ,y p (x) represents the coordinates of the center point of the prediction box. g ,y g The coordinates of the center point of the ground truth bounding box are shown below. The coordinates of the center point of the predicted bounding box are shown below. coordinates of the center point of the true bounding box Substituting into the distance formula, we can obtain

[0106] Calculate the diagonal length c(B) of the minimum closure region covering the predicted bounding box and the ground truth bounding box. p B gSpecifically, this includes calculating the difference between the top-left and bottom-right coordinates of the two bounding boxes. For example, if the top-left corner coordinate of the predicted bounding box is (x... p1 ,y p1 The coordinates of the lower right corner are (x p2 ,y p2 The coordinates of the top-left corner of the true bounding box are (x... g1 ,y g1 The coordinates of the lower right corner are (x g2 ,y g2 ), then c(B) p B g )=max(|x p2 -x g1 |,|x g2 -x p1 |)+max(|y p2 -y g1 |,|y g2 -y p1 |)=max(|55-20|,|60-25|)+max(|75-30|,|80-35|)=35+45=80.

[0107] Next, define the target size function s(B) to determine the target size based on factors such as the area of ​​the target bounding box. Where (x) p1 ,y p1 ) and (x p2 ,y p2 () represents the diagonal coordinates of the prediction box; Where (x) g1 ,y g1 ) and (x g2 ,y g2 () represents the diagonal coordinates of the ground truth bounding box. For the predicted bounding box area... For the true frame area

[0108] Therefore, it can be calculated that The scale-adaptive cross-union ratio loss function is established as follows: The parameter μ has a value range of -∞ < α < 1. In this embodiment, μ = -3, and its function is to adjust the scaling level for small targets. The parameter ν is the rate at which larger objects are restored to the standard intersection-union ratio. Its value range is 1 < ν < ∞. In this embodiment, ν = 16. By adjusting the values ​​of μ and v, the network can adaptively adjust the detection accuracy for targets of different scales.

[0109] Through the complete data flow and specific numerical calculations described above, the calculation process of each part in adaptive feature extraction, multi-scale enhanced feature pyramid network, and scale-adaptive intersection-union loss function, as well as their overall application, are demonstrated. In the actual training process of object detection networks, such loss calculations are performed based on a large amount of sample bounding boxes (real and predicted boxes). Optimization algorithms (such as gradient descent) are then used to adjust the network parameters based on the loss value, enabling the network to adaptively adjust the detection accuracy for both small and large targets, thereby improving the overall object detection performance.

[0110] In one embodiment, the target detection method based on adaptive feature extraction and multi-scale enhancement may further include a process of outputting target detection results. The specific process includes: inputting deep features into a classifier and a regressor; using the classifier to predict the target category based on the deep features to obtain the probability distribution of the target belonging to different categories; using the regressor to predict the target position and size to obtain the position and size prediction results; and outputting the probability distribution, position and size prediction results as the target detection results.

[0111] That is, the fused features are processed by classifiers and regressors to output the final target detection results, including information such as the target's category, location, and size.

[0112] The classifier employs structures such as fully connected layers and a softmax function to predict the class of the feature map, outputting the probability distribution of each object belonging to different classes. For example, for an object detection task with N classes, the classifier outputs an N-dimensional probability vector, representing the likelihood of an object belonging to each class.

[0113] The regressor uses structures such as convolutional layers or fully connected layers to predict the location and size of targets on the feature map. For example, it predicts the center coordinates (x, y) of the target detection box, as well as parameters such as width w and height h.

[0114] Based on the outputs of the classifier and regressor, the final target detection result is determined, including information such as the target's category, location, and size. For example, the category with the highest classification probability is selected as the target category, and the target's location and size in the image are determined based on the coordinates and dimensions predicted by the regressor. This information is then compiled into a list of target detection results for output.

[0115] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0116] In one embodiment, such as Figure 5 As shown, a target detection system based on adaptive feature extraction and multi-scale enhancement is provided, including: an image processing module 510, an adaptive feature extraction module 520, a multi-scale feature enhancement module 530, and a target detection module 540, wherein:

[0117] Image processing module 510 is used to acquire the input raw image and perform standardization preprocessing on the raw image to obtain the raw image to be detected;

[0118] The adaptive feature extraction module 520 is used to input the original image to be detected into the dilated convolution block of the parallel structure, and capture the feature information of the original image to be detected at different scales through the convolutional layers with different dilation rates in the dilated convolution block, extract the image features, and obtain the feature maps of each channel.

[0119] The adaptive feature extraction module 520 is also used to determine the target normalization method based on image features. The contribution of each channel feature is dynamically adjusted through the target normalization method to obtain the optimal feature vector. The optimal feature vector is scaled to obtain the compressed target attention weight. The channel attention gating method is used to calculate the gating signal for the target attention weight, and a fused feature map is generated based on the gating signal and the channel feature map.

[0120] The multi-scale feature enhancement module 530 is used to input the fused feature map into the multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the middle layer, thereby achieving multi-scale feature fusion and extracting deep features.

[0121] The target detection module 540 is used to input deep features into the classifier and regressor, dynamically adjust the cross-union ratio threshold using the scale-adaptive cross-union ratio loss function, and output the target detection results.

[0122] In one embodiment, the image processing module 510 is further configured to acquire the input original image and acquire the pixel matrix of the original image; and to normalize the pixel matrix according to a predetermined data standardization rule to obtain the original image to be detected.

[0123] In one embodiment, the adaptive feature extraction module 520 is further configured to input the original image to be detected into the dilated convolution block of the parallel structure in the form of tensors; perform parallel feature extraction through convolutional layers with different dilation rates in the dilated convolution block to obtain feature maps at different scales and receptive fields; and fuse the feature maps in the channel dimension by feature concatenation to obtain the fused channel feature maps.

[0124] In one embodiment, the adaptive feature extraction module 520 is further configured to perform spatial dimension normalization processing on the feature maps of each channel using various normalization methods to obtain normalization processing results; select the best normalization method as the target normalization method based on the normalization processing results using an activation function; and dynamically adjust the contribution of each channel feature using the target normalization method to obtain the optimal feature vector.

[0125] In one embodiment, the adaptive feature extraction module 520 is further configured to scale the optimal feature vector using a set of learnable weight parameters and to scale it using a parameterless normalization technique to obtain a compressed feature vector; to introduce a specific scalar into the compressed feature vector to normalize the scale, and to perform scale normalization to obtain the compressed target attention weights.

[0126] In one embodiment, the adaptive feature extraction module 520 is further configured to use an activation function to apply a channel attention gating method to the target attention weights to calculate a gate control signal; use the target attention weights to weight the channel feature map to obtain a weighted feature map; multiply the gate control signal and the weighted feature map element by element to obtain the result of the multiplication; and sum the result of the multiplication along the first dimension to obtain a fused feature map.

[0127] In one embodiment, the structure of the adaptive feature extraction module is as follows: Figure 6As shown, the main components include a parallel-structured dilated convolution module, adaptive normalization, activation functions, and a channel attention gating module. Specifically: the parallel-structured dilated convolution module extracts features from the preprocessed image data using a progressively dilated convolution sequence and concatenates them along the channel dimension; then, it applies an adaptive normalization method to learn the importance of channel feature maps in different branches, dynamically adjusting the contribution of each channel feature to enhance the model's sensitivity to key features; the channel attention gating module uses a standard normalization library, adaptively selecting the optimal normalization method based on the input data; and finally, it uses the softmax activation function to adaptively select the optimal normalization method, obtaining the best feature vector.

[0128] In one embodiment, the multi-scale feature enhancement module 530 is further configured to input the fused feature map into a multi-scale enhanced feature pyramid network, scale the features in the fused feature map through a linear scaling layer in the multi-scale enhanced feature pyramid network, perform pixel-by-pixel addition on the scaled features and the original features, and then input them into the fusion layer in the multi-scale enhanced feature pyramid network; perform convolution operation on the input features through the fusion layer to extract and achieve multi-scale feature fusion, thereby obtaining the deep features of the fused feature map.

[0129] In one embodiment, the target detection system based on adaptive feature extraction and multi-scale enhancement may further include a scale-adaptive cross-union ratio (CURRR) loss function module, used to determine the ground truth bounding box and the predicted bounding box of the target in the original image to be detected, and to calculate the CURRR between the ground truth bounding box and the predicted bounding box; calculate the Euclidean distance between the center points of the ground truth bounding box and the predicted bounding box, and calculate the diagonal length of the minimum closure region covering the ground truth bounding box and the predicted bounding box; define a target size function based on the diagonal length, and obtain the scale-adaptive CURRR loss function according to the CURRR, the Euclidean distance, the diagonal length, and the target size function.

[0130] In one embodiment, the target detection module 540 is further configured to input deep features into a classifier and a regressor, and the classifier predicts the target category based on the deep features to obtain the probability distribution of the target belonging to different categories; the regressor predicts the target position and size to obtain the position and size prediction results; and the probability distribution, position and size prediction results are output as the target detection results.

[0131] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a target detection method based on adaptive feature extraction and multi-scale enhancement. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0132] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0133] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a target detection method based on adaptive feature extraction and multi-scale enhancement.

[0134] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a target detection method based on adaptive feature extraction and multi-scale enhancement.

[0135] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0136] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0137] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A target detection method based on adaptive feature extraction and multi-scale enhancement, characterized in that, The method includes: The original input image is acquired, and the original image is subjected to standardization preprocessing to obtain the original image to be detected; The original image to be detected is input into a parallel structured dilated convolutional block. The convolutional layers with different dilation rates in the dilated convolutional block capture the feature information of the original image to be detected at different scales, extract the image features, and obtain feature maps for each channel. Based on the image features, a target normalization method is determined. The contribution of each channel feature is dynamically adjusted through the target normalization method to obtain the optimal feature vector. The optimal feature vector is then scaled to obtain the compressed target attention weight. A channel attention gating method is used to calculate the gating signal for the target attention weight, and a fused feature map is generated based on the gating signal and the channel feature map. The fused feature map is input into a multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the middle layer, thereby achieving multi-scale feature fusion and extracting deep features. The deep features are input into the classifier and regressor, and the cross-union ratio threshold is dynamically adjusted using the scale-adaptive cross-union ratio loss function to output the target detection results. The method further includes: determining the ground truth bounding box and the predicted bounding box of the target in the original image to be detected, and calculating the intersection-union ratio (IU) of the ground truth bounding box and the predicted bounding box; calculating the Euclidean distance between the center points of the ground truth bounding box and the predicted bounding box, and calculating the diagonal length of the minimum closure region covering the ground truth bounding box and the predicted bounding box; defining a target size function based on the diagonal length, and obtaining a scale-adaptive IU loss function based on the IU, the Euclidean distance, the diagonal length, and the target size function.

2. The target detection method based on adaptive feature extraction and multi-scale enhancement according to claim 1, characterized in that, The process involves acquiring the input raw image and performing a standardization preprocessing on the raw image to obtain the raw image to be detected, including: Obtain the input original image and obtain the pixel matrix of the original image; The pixel matrix is ​​normalized according to predetermined data standardization rules to obtain the original image to be detected.

3. The target detection method based on adaptive feature extraction and multi-scale enhancement according to claim 1, characterized in that, The original image to be detected is input into a parallel dilated convolutional block. Convolutional layers with different dilation rates within the dilated convolutional block capture feature information of the original image at different scales, extracting image features to obtain feature maps for each channel, including: The original image to be detected is input as a tensor into a dilated convolution block of a parallel structure; Parallel feature extraction is performed using convolutional layers with different dilation rates in the dilated convolutional block to obtain feature maps at various scales and receptive fields. The feature maps are fused along the channel dimension using a feature concatenation method to obtain fused channel feature maps.

4. The target detection method based on adaptive feature extraction and multi-scale enhancement according to claim 1, characterized in that, Based on the image features, a target normalization method is determined. This method dynamically adjusts the contribution of each channel's features to obtain the optimal feature vector, including: Each of the channel feature maps is normalized in spatial dimension using various normalization methods to obtain the normalization result; The optimal normalization method is selected as the target normalization method based on the normalization result using the activation function. The contribution of each channel feature is dynamically adjusted using the target normalization method to obtain the optimal feature vector.

5. The target detection method based on adaptive feature extraction and multi-scale enhancement according to claim 4, characterized in that, The optimal feature vector is scaled to obtain compressed target attention weights, including: The optimal feature vector is scaled using a set of learnable weight parameters and then scaled again using a parameterless normalization technique to obtain a compressed feature vector. A specific scalar is introduced into the compressed feature vector to normalize the scale, and scale normalization is performed to obtain the compressed target attention weights.

6. The target detection method based on adaptive feature extraction and multi-scale enhancement according to claim 5, characterized in that, A channel attention gating method is used to calculate a gating signal for the target attention weights, and a fused feature map is generated based on the gating signal and the channel feature map, including: The target attention weights are applied to a channel attention gating method using an activation function to calculate the gate control signal. The channel feature map is weighted using the target attention weights to obtain a weighted feature map; The gated signal is multiplied element by element by the weighted feature map to obtain the result of the multiplication; The result of the multiplication is summed along the first dimension to obtain the fused feature map.

7. The target detection method based on adaptive feature extraction and multi-scale enhancement according to claim 1, characterized in that, The fused feature map is input into a multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the intermediate layer, achieving multi-scale feature fusion and extracting deep features, including: The fused feature map is input into a multi-scale augmented feature pyramid network, and the features in the fused feature map are scaled by a linear scaling layer in the multi-scale augmented feature pyramid network. The scaled features are then added pixel-by-pixel to the original features and input into the fusion layer of the multi-scale enhanced feature pyramid network. The fusion layer performs convolution operations on the input features to extract and achieve multi-scale feature fusion, thereby obtaining the deep features of the fused feature map.

8. The target detection method based on adaptive feature extraction and multi-scale enhancement according to claim 1, characterized in that, The deep features are input into the classifier and regressor, and the cross-union ratio (CUI) threshold is dynamically adjusted using a scale-adaptive CUI loss function. The target detection results are then output, including: The deep features are input into the classifier and regressor. The classifier predicts the target category based on the deep features to obtain the probability distribution of the target belonging to different categories. The target position and size are predicted using the regressor, and the prediction results are obtained. The probability distribution, location, and size prediction results are output as target detection results.

9. A target detection system based on adaptive feature extraction and multi-scale enhancement, characterized in that, The system includes: The image processing module is used to acquire the input raw image and perform standardization preprocessing on the raw image to obtain the raw image to be detected; An adaptive feature extraction module is used to input the original image to be detected into a parallel structured dilated convolution block, and capture the feature information of the original image to be detected at different scales through convolutional layers with different dilation rates in the dilated convolution block, extract image features, and obtain feature maps of each channel. The adaptive feature extraction module is further configured to determine a target normalization method based on the image features, dynamically adjust the contribution of each channel feature through the target normalization method to obtain the optimal feature vector, scale the optimal feature vector to obtain the compressed target attention weight, calculate the gating signal for the target attention weight using the channel attention gating method, and generate a fused feature map based on the gating signal and the channel feature map. The multi-scale feature enhancement module is used to input the fused feature map into the multi-scale enhanced feature pyramid network. Based on the fused feature map, the multi-scale enhanced feature pyramid network scales the upper and lower layer feature maps to match the scale of the middle layer, thereby achieving multi-scale feature fusion and extracting deep features. The target detection module is used to input the deep features into the classifier and regressor, dynamically adjust the cross-union ratio threshold using the scale-adaptive cross-union ratio loss function, and output the target detection results. It also includes a scale-adaptive cross-union ratio (CURRR) loss function module, which is used to determine the ground truth bounding box and the predicted bounding box of the target in the original image to be detected, and calculate the CURRR between the ground truth bounding box and the predicted bounding box; calculate the Euclidean distance between the center points of the ground truth bounding box and the predicted bounding box, and calculate the diagonal length of the minimum closure region covering the ground truth bounding box and the predicted bounding box; define the target size function based on the diagonal length, and obtain the scale-adaptive CURRR loss function according to the CURRR, Euclidean distance, diagonal length, and target size function.

Images