Explosive object target identification method based on multi-dimensional data optimization and feature fusion

Abstract

The invention relates to the technical field of computer vision and artificial intelligence image processing, and discloses an explosive object target identification method based on multi-dimensional data optimization and feature fusion, and the method comprises the following steps: carrying out the HSV space brightness adjustment, random cutting and gray conversion of an original image, and generating a single-channel gray image; extracting a multi-level feature map by using a convolutional neural network; utilizing a fixed gradient operator to generate a structure attention mask enhanced superficial layer feature map; re-standardizing each feature map to generate a multi-scale feature map with aligned distribution; a bidirectional cross-scale transmission path is constructed and weighted fusion is carried out; and performing category prediction and bounding box regression through the decoupling prediction network, and outputting an explosive object detection result. According to the method, random disturbance and single-channel graying preprocessing logic based on an HSV brightness channel is constructed, non-essential color interference can be actively stripped, a model is forced to focus on inherent geometric texture features of an object, and therefore the illumination invariance basis is established at the data input end.

Description

Technical Field

[0001] This invention relates to the field of computer vision and artificial intelligence image processing technology, specifically to a method for explosive target recognition based on multi-dimensional data optimization and feature fusion. Background Technology

[0002] In mining and geological exploration operations, the control of the flow of explosives, detonators, and other civilian explosive materials is directly related to production safety and public security. The work sites are often in harsh environments, and how to confirm the real-time status of these hazardous materials and prevent their loss or unauthorized removal is a core challenge for on-site safety supervision. Traditional management methods rely heavily on manual on-site inventory or visual review of surveillance videos. This approach is not only inefficient but also prone to loopholes due to human fatigue or negligence. Therefore, utilizing computer vision technology for automated real-time detection and identification of explosive materials in on-site video streams has become a key means to improve the intelligence level of hazardous explosives management and reduce safety risks.

[0003] Existing general-purpose object detection technologies primarily rely on deep convolutional neural networks. Through training on large-scale datasets, these models can automatically learn feature representations from images. Algorithms such as the R-CNN series or YOLO series have demonstrated high accuracy in urban security or traffic monitoring fields with ample lighting and clear scene structures. These algorithms, through deep network stacking, can effectively extract the color texture and high-level semantic information of objects. When processing conventional objects with obvious appearance features, they possess rapid response and classification capabilities, greatly promoting the engineering application of image recognition technology.

[0004] However, directly transplanting the aforementioned general technologies to underground or field blasting scenarios reveals their limitations. First, existing algorithms rely excessively on RGB color information, while underground mines are mostly illuminated by moving point light sources with drastic fluctuations in light intensity and large color temperature deviations. This lighting noise causes the color distribution of object imaging to drift, rendering the model's original color features ineffective. Second, existing networks tend to capture obvious texture blocks, but targets such as detonators and fuses are small in size and often covered by dust, resulting in extremely low contrast with the background rocks. Networks lacking prior guidance on physical edges struggle to separate these subtle geometric contours from the murky background, easily leading to missed detections. Furthermore, when handling multi-scale feature fusion, general algorithms often employ a simple linear superposition method. When simultaneously detecting large packaging boxes and small detonators, shallow background noise easily overwhelms deep, weak semantic information, resulting in the positioning accuracy of small-scale targets failing to meet the stringent requirements of high-risk goods control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for identifying explosive targets based on multi-dimensional data optimization and feature fusion. This method aims to solve the problems of poor generalization robustness caused by over-reliance on RGB color distribution in complex lighting scenarios such as underground mines, as well as the difficulty in extracting geometric edge features of small explosives and insufficient positioning accuracy under low-contrast backgrounds and large-scale spans.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying explosive targets based on multi-dimensional data optimization and feature fusion, comprising the following steps: The original image to be identified is acquired, and HSV spatial brightness adjustment, random cropping and grayscale conversion are performed on the original image to generate a single-channel grayscale image. A convolutional neural network is used to perform multi-level feature extraction on the single-channel grayscale image, outputting shallow feature maps, mid-level feature maps, and deep feature maps; Edge information of the shallow feature map is extracted using a fixed gradient operator to generate a structural attention mask, and the structural attention mask is applied to the shallow feature map to obtain an enhanced shallow feature map. The spatial dimension statistics of the enhanced shallow feature map, the middle feature map, and the deep feature map are calculated and re-standardized to generate a multi-scale feature map with distribution alignment. The channel dimension of the distributed and aligned multi-scale feature maps is unified, a bidirectional cross-scale transmission path is constructed and weighted fusion is performed to generate a fused multi-scale feature map. The decoupled prediction network is used to perform category prediction and bounding box regression based on the fused multi-scale feature map, and the explosive body detection results are output.

[0007] Preferably, the step of performing HSV spatial brightness adjustment, random cropping, and grayscale conversion on the original image to generate a single-channel grayscale image includes: The original image is converted from the RGB color space to the HSV color space, keeping the hue and saturation components unchanged, and only applying a linear transformation to the lightness component; The adjusted brightness component is merged with the original hue and saturation components and then inversely transformed back to the RGB color space. Read the annotation file of the original image, randomly generate a cropping window under the constraint of satisfying a preset coverage ratio threshold, and scale the cropping area to a preset network input size; The weighted average method is used to calculate the weighted sum of the pixel values ​​of each RGB channel, and the color image is converted into the single-channel grayscale image.

[0008] Preferably, the step of using a convolutional neural network to perform multi-level feature extraction on the single-channel grayscale image and outputting shallow feature maps, mid-level feature maps, and deep feature maps includes: Construct a convolutional neural network that includes depthwise separable convolution and channel attention mechanisms, and adapt the input channel parameters of the first convolutional layer to a single channel; The weights of the first-layer convolutional kernel are initialized using the mean weight transfer strategy to keep the amplitude distribution of the convolutional kernel's response to the single-channel input signal approximately consistent with the pre-training state. Through layer-by-layer downsampling operations, the shallow feature map with a downsampling factor of 8, the middle feature map with a downsampling factor of 16, and the deep feature map with a downsampling factor of 32 are output respectively.

[0009] Preferably, the step of extracting edge information from the shallow feature map using a fixed gradient operator to generate a structural attention mask, and applying the structural attention mask to the shallow feature map to obtain an enhanced shallow feature map, includes: Using preset, non-learnable horizontal and vertical convolutional kernels, depthwise convolution operations are performed on the shallow feature maps respectively, outputting horizontal gradient feature maps and vertical gradient feature maps; Calculate the combined gradient magnitude of the horizontal gradient feature map and the vertical gradient feature map at corresponding spatial locations; The gradient synthesis magnitude is renormalized by applying learnable affine transformation parameters and mapped to the structure attention mask with values ​​between 0 and 1 via a nonlinear activation function. The structural attention mask is multiplied element-wise with the shallow feature map, and the result is added to the original shallow feature map to obtain the enhanced shallow feature map.

[0010] Preferably, the step of calculating the spatial dimension statistics of the enhanced shallow feature map, the middle feature map, and the deep feature map respectively, and then re-standardizing them to generate a distribution-aligned multi-scale feature map includes: An instance normalization layer is constructed independently for the feature map of each level; For each sample and each channel in the feature map, calculate the pixel mean and pixel variance of the channel in the spatial dimension. The feature pixel values ​​of the channel are standardized with zero mean and unit variance using the pixel mean and pixel variance. The standardized features are subjected to affine transformation using learnable scaling and translation factors to output the distribution-aligned multi-scale feature map.

[0011] Preferably, the process of unifying the channel dimensions of the distributed, aligned multi-scale feature maps, constructing a bidirectional cross-scale transfer path, and performing weighted fusion to generate the fused multi-scale feature map includes: The input feature maps at each level are convolved point by point using 1×1 convolution kernels, and the number of channels in all feature maps is adjusted to the preset dimension. Remove intermediate nodes in the merged topology that have only a single input edge; Construct a top-down semantic enhancement path, and pass high-level features to low-level features after upsampling; Construct a bottom-up localization enhancement path, which downsamples low-level features and then passes them to high-level features; Establish direct jump connections between input and output nodes at the same resolution level; For each fusion node at each level, the input features from different paths are normalized and weighted by using learnable weight parameters.

[0012] Preferably, the step of using a decoupled prediction network to perform category prediction and bounding box regression based on the fused multi-scale feature map, and outputting explosive body detection results, includes: Construct a parameter-sharing classification subnetwork and a box regression subnetwork, both of which are composed of stacked depthwise separable convolutional layers; The classification subnetwork outputs logical values ​​indicating that each preset anchor box belongs to a specific category, and the activation function maps the logical values ​​to category confidence. The box regression subnetwork outputs the positional offset and size scaling factor of the predicted box relative to the preset anchor box; Based on the downsampling factor of the feature map and the geometric parameters of the preset anchor box, the absolute center coordinates and width and height dimensions of the prediction box in the original image are calculated using the position offset and the size scaling factor. Non-maximum suppression is applied to all predicted boxes based on the cross-union ratio threshold to remove redundant overlapping detection boxes.

[0013] Preferably, applying a linear transformation only to the lightness component includes: Generate random brightness adjustment coefficients that follow a uniform distribution; The original luminance component is linearly gained using the random luminance adjustment coefficient, and the calculation result is restricted to the minimum and maximum pixel values ​​allowed by the image format using a numerical truncation function. The random generation of the cropping window under the constraint of satisfying a preset coverage ratio threshold includes: Calculate the area of ​​the intersection between the randomly generated cropping window region and the target true bounding box region; Determine whether the ratio of the intersection area to the target true bounding box region is greater than or equal to a preset coverage ratio threshold. If the ratio is satisfied, retain the clipping window.

[0014] Preferably, the step of performing depthwise convolution operations on the shallow feature map using preset, non-learnable horizontal and vertical convolution kernels includes: The weight matrix of the horizontal convolution kernel is set to the horizontal operator value of the Sobel operator, and the weight matrix of the vertical convolution kernel is set to the vertical operator value of the Sobel operator. The weights of the horizontal and vertical convolutional kernels are not updated during training. The step of renormalizing the gradient synthesis magnitude by applying learnable affine transformation parameters and mapping it via a nonlinear activation function to the structure attention mask with values ​​between 0 and 1 includes: Multiply the gradient composite magnitude by the learnable scaling factor and then add the learnable bias term; The weighted result is input into the Sigmoid function, so that the gradient magnitude of the background region is mapped to a range close to 0, and the gradient magnitude of the edge region is mapped to a range close to 1.

[0015] Preferably, the step of normalizing and weighting the input features from different paths using learnable weight parameters for each fusion node includes: For each input feature, the ReLU activation function is applied to the learnable weight parameter to force the weight parameter to be non-negative. Calculate the sum of the products of each input feature and its corresponding weight parameter; Calculate the sum of all the weight parameters and add the numerical stability constant to the sum; Divide the sum of the products by the sum after adding the numerical stability constant to obtain the fused feature output.

[0016] This invention provides a method for identifying explosive targets based on multi-dimensional data optimization and feature fusion. It has the following beneficial effects: 1. This invention constructs a preprocessing logic based on random perturbation and single-channel grayscale conversion of the HSV brightness channel, which can actively remove non-essential color interference and force the model to focus on the inherent geometric texture features of the object. This establishes the basis of illumination invariance at the data input end. Compared with the traditional recognition scheme that relies on RGB three-channel input, this mechanism effectively avoids color temperature shift and overexposure interference caused by complex lighting conditions in the well, and significantly improves the generalization robustness of the model under extreme lighting conditions such as backlight and darkness.

[0017] 2. This invention introduces a structured attention module with an embedded fixed gradient operator. It guides shallow feature extraction through physical edge priors and uses learnable affine transformations to renormalize the gradient magnitude to generate a high-response mask, thereby achieving targeted enhancement of high-frequency geometric information. In response to the shortcomings of existing pure data-driven convolutional networks that are prone to overfitting to background noise in low-contrast scenes, this scheme enhances the feature response of fine structures such as detonator leads and packaging creases through hard inductive bias, effectively solving the problem of feature submersion and missed detection of blurred edge targets.

[0018] 3. This invention implements a feature distribution alignment and bidirectional weighted feature pyramid fusion strategy based on instance normalization. While eliminating the statistical distribution shift caused by strong data augmentation, it uses fast normalized weights to achieve nonlinear adaptive aggregation of cross-scale information. Unlike the equal-weight direct superposition method used by conventional feature pyramid networks, this technical solution can dynamically allocate weights according to the actual contribution of different levels of features to the recognition task. It solves the problem of information mutual exclusion and dilution between deep semantics and shallow details in the fusion process, and greatly improves the positioning accuracy of multi-scale aliasing and small target explosive bodies. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of the explosive target recognition system based on multi-dimensional data optimization and feature fusion according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the explosive target recognition method based on multi-dimensional data optimization and feature fusion according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the multi-dimensional data optimization preprocessing process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the logical operation flow of the gradient-guided structured attention module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structural topology and data flow of the bidirectional feature pyramid fusion module in an embodiment of the present invention.

[0020] in: 10. Multi-dimensional data optimization preprocessing module; 20. Feature extraction backbone network module; 30. Gradient-guided structured attention module; 40. Brightness-invariant feature alignment module; 50. Bidirectional feature pyramid fusion module; 60. Target detection output module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 This invention provides an explosive target recognition system based on multi-dimensional data optimization and feature fusion. The system includes a multi-dimensional data optimization preprocessing module 10, a feature extraction backbone network module 20, a gradient-guided structured attention module 30, a brightness-invariant feature alignment module 40, a bidirectional feature pyramid fusion module 50, and a target detection output module 60.

[0023] The multi-dimensional data optimization preprocessing module 10 is used to standardize and enhance the raw acquired image data to generate input data suitable for model training or inference. The multi-dimensional data optimization preprocessing module 10 is configured with HSV-based random brightness adjustment logic to convert the original image from RGB space to HSV space and apply a linear transformation to the V channel to simulate changes in illumination intensity. The multi-dimensional data optimization preprocessing module 10 is also configured with bounding box-based random cropping logic to generate cropping regions according to preset area coverage ratio constraints. The multi-dimensional data optimization preprocessing module 10 is further configured with grayscale conversion logic to convert the brightness-adjusted and cropped image into a single-channel grayscale image. The output of the multi-dimensional data optimization preprocessing module 10 is connected to the input of the feature extraction backbone network module 20.

[0024] The feature extraction backbone network module 20 is used to perform multi-level convolutional feature extraction on the input single-channel grayscale image. The feature extraction backbone network module 20 is built based on a convolutional neural network architecture and generates feature maps of different resolutions through layer-by-layer downsampling operations. The feature extraction backbone network module 20 outputs feature maps at least three levels: a shallow feature map containing high-resolution geometric information, a mid-level feature map containing medium-level semantic information, and a deep feature map containing high-level semantic information.

[0025] The gradient-guided structured attention module 30 is connected to the port of the feature extraction backbone network module 20 that outputs the shallow feature map. The gradient-guided structured attention module 30 is used to extract edge features from the shallow feature map using a differentiable gradient operator. The gradient-guided structured attention module 30 calculates the gradient magnitudes of the shallow feature map in the horizontal and vertical directions and generates a structured attention mask based on the gradient magnitudes. The gradient-guided structured attention module 30 performs element-wise operations between the structured attention mask and the original shallow feature map, outputting a geometrically enhanced shallow feature map.

[0026] The brightness-invariant feature alignment module 40 is connected to the output of the gradient-guided structured attention module 30 and the outputs of the mid-level and deep feature maps of the feature extraction backbone network module 20. The brightness-invariant feature alignment module 40 is used to eliminate the distribution shift in the feature maps caused by random brightness adjustment in the multi-dimensional data optimization preprocessing module 10. The brightness-invariant feature alignment module 40 performs instance normalization calculations on the input feature maps at each scale, standardizing the feature distribution by independently calculating the statistics of each sample and each channel.

[0027] The bidirectional feature pyramid fusion module 50 is connected to the output of the brightness-invariant feature alignment module 40. The bidirectional feature pyramid fusion module 50 is used for cross-scale information interaction of the normalized multi-scale feature maps. The bidirectional feature pyramid fusion module 50 constructs a top-down semantic feature transfer path and a bottom-up localization feature transfer path. The bidirectional feature pyramid fusion module 50 configures direct skip connections between input and output nodes at the same resolution level. The bidirectional feature pyramid fusion module 50 is configured with fast normalized weighted fusion logic, which uses learnable weight parameters to perform weighted summation of multiple input features.

[0028] The target detection output module 60 is connected to the output of the bidirectional feature pyramid fusion module 50. The target detection output module 60 is used for regression and classification calculations based on the fused multi-scale feature map. The target detection output module 60 outputs the class confidence score of the explosive and the coordinate information of the bounding box surrounding the target object.

[0029] See attached document Figure 2 This invention provides a method for identifying explosive targets based on multi-dimensional data optimization and feature fusion, comprising the following steps: S100: The original image to be identified is obtained through the multi-dimensional data optimization preprocessing module 10. The original image is converted to the HSV color space and the V channel is separated and a random brightness coefficient is applied for adjustment. A random cropping box that meets the preset area ratio is generated according to the target annotation information. The cropped image area is converted into a single-channel grayscale image. S200, the feature extraction backbone network module 20 receives a single-channel grayscale image as an input tensor, performs multi-level convolution and pooling operations, and outputs shallow feature maps, medium feature maps and deep feature maps. S300, the gradient-guided structured attention module 30 receives the shallow feature map, performs convolution operation on the shallow feature map using the Sobel operator to obtain the gradient map, generates the attention weight matrix through nonlinear activation, and applies the attention weight matrix to the shallow feature map to obtain the enhanced shallow feature map. S400, the brightness invariance feature alignment module 40 receives the enhanced shallow feature map, middle feature map and deep feature map respectively, calculates the mean and variance of each feature map in the spatial dimension, and re-standardizes the pixel values ​​of the feature map according to the calculation results, and outputs a multi-scale feature map with distribution alignment. S500, the bidirectional feature pyramid fusion module 50 receives multi-scale feature maps with distribution alignment, adjusts the channel dimension of each feature map to a uniform value, performs iterative bidirectional feature fusion operation, and integrates feature information from different levels through weighted accumulation. S600, the target detection output module 60 receives the fused feature map, calculates the target category probability and position offset through the convolutional prediction layer, and outputs the explosive body detection result.

[0030] See attached document Figure 3 The multi-dimensional data optimization preprocessing step S100 aims to construct data inputs that adapt to complex lighting and scale changes. In this embodiment, to improve the model's robustness to complex downhole environments, this preprocessing is mainly applied to the data augmentation stage during model training; during the model inference application stage, only necessary format conversions and size scaling can be performed according to actual needs. Specifically, it includes the following sub-steps: S101, perform random brightness adjustment based on the HSV color space. In this embodiment, the multi-dimensional data optimization preprocessing module 10 first receives the original acquired RGB format image, and uses a color space conversion algorithm to convert the RGB image to the HSV color space, obtaining image data containing hue component H, saturation component S, and lightness component V. During this process, the values ​​of hue component H and saturation component S are kept unchanged, and only the lightness component V is operated on. The technical principle is to separate the change in light intensity from the inherent color characteristics of the object, avoiding color distortion caused by directly adjusting the RGB channels.

[0031] The system generates a random brightness adjustment coefficient that follows a uniform distribution. The coefficient's value range is set to [-0.3, 0.3]. This range is determined based on statistical data of illumination fluctuations in actual mines and field blasting scenarios. Used to simulate backlighting or shadow scenes Used to simulate scenes under strong light. Utilizes random brightness adjustment coefficients. For the original lightness component Perform a linear transformation and calculate the new lightness components. The brightness adjustment formula is as follows: ; in, Indicates the adjusted pixel coordinates The brightness value at that location, Indicates the pixel coordinates before adjustment The original brightness value at that location, This is a randomly generated brightness adjustment coefficient. This represents a numerical truncation function. and These are the minimum and maximum pixel values ​​allowed by the image format (0 and 255 respectively for the 8-bit image in this embodiment), used to prevent image artifacts caused by numerical overflow after brightness enhancement.

[0032] After adjusting the brightness channel, the multi-dimensional data optimization preprocessing module 10 will process the adjusted brightness components. The color image is then merged with the original hue component H and saturation component S, and an inverse HSV to RGB transformation algorithm is used to generate a color image with enhanced brightness. The specific matrix operations for color space conversion and inverse transformation are well-known techniques in the field of image processing and can be directly implemented by those skilled in the art, so they will not be described in detail here.

[0033] S102, perform random cropping based on target bounding box constraints. During the training data construction phase, the multi-dimensional data optimization preprocessing module 10 reads the annotation file corresponding to the original image to obtain the true bounding box of the explosive body in the image, denoted as... To enhance the model's robustness to changes in target scale and local occlusion, the system randomly generates a cropping window within a 10-pixel effective region after removing image edges. The operation of removing edge pixels is to eliminate edge black border interference caused by lens distortion or image acquisition equipment. The generation of this cropping window must meet preset physical constraints, that is, the area of ​​the target pixels contained within the cropping area must be greater than a preset coverage ratio threshold of the total target area.

[0034] In this embodiment, a preset coverage ratio threshold is used. The threshold is set to 0.7. This threshold is chosen because the identification of target objects such as detonators and explosive charges relies on the combined features of the fuse, tube body, and outer packaging texture. When the retained area is less than 70%, key combined features may be completely lost, leading to semantic ambiguity in the generated training samples. Conversely, a local view with more than 70% retention can effectively simulate foreground occlusion, prompting the network to learn local discriminative features. The formula for determining the pruning constraint is as follows: ; in, This refers to the operation used to calculate the area of ​​a region. This represents a randomly generated cropping window area. This represents the true bounding box region of the target. This represents the operation of finding the intersection of two regions. This represents a preset coverage ratio threshold, in this embodiment... .

[0035] If the randomly generated clipping window does not meet the above constraints, the system will regenerate it until the conditions are met. After obtaining the valid clipping region, the system uses bilinear interpolation to uniformly scale the region to the preset network input size. (In this embodiment, it is set to 640×640 pixels).

[0036] S103, performs grayscale space conversion and desaturation. The multi-dimensional data optimization preprocessing module 10 receives the color image after brightness adjustment and cropping / scaling, and converts it into a single-channel grayscale image using a weighted average method. The technical objective of this step is to reduce the dimensionality of the input data and actively remove color features. Because the outer packaging color of explosives (such as fuses and emulsion explosives) varies drastically between different batches and under different lighting temperatures, color features often constitute misleading noise. By using grayscale processing, the subsequent network no longer relies on unstable color information, but instead focuses on more stable geometric contours and texture gradient features.

[0037] The formula for grayscale conversion is as follows: ; in, This indicates the output single-channel grayscale image in coordinates. Pixel value at that location, , , These represent the pixel values ​​of the input color image in the red, green, and blue channels, respectively. , , These are the grayscale weight coefficients for each color channel, which are set to values ​​of [value] in this embodiment. , , This weighting coefficient reflects the contribution of different color channels to brightness information, ensuring that the converted grayscale image can retain the texture contrast of the original image to the greatest extent.

[0038] The single-channel grayscale image generated after processing in steps S101 to S103 retains the enhanced geometric texture information and eliminates color noise, serving as the standard input data for the subsequent feature extraction backbone network module 20.

[0039] Step S200, which involves constructing the backbone network for feature extraction, aims to establish a multi-level deep convolutional neural network architecture. This architecture transforms the input single-channel grayscale image into a set of feature pyramids with different resolutions, channel dimensions, and semantic levels, providing foundational feature maps for subsequent feature enhancement and fusion. This step specifically includes the following sub-steps: S201, Construct a convolutional neural network architecture adapted to single-channel input. In this embodiment, the feature extraction backbone network module 20 uses EfficientNet-B0 as its basic architecture. The core component of this architecture is the Moving Inverted Bottleneck Convolutional Module (MBConv), which integrates depthwise separable convolution and channel attention mechanisms. The technical advantage of this architecture is that by decomposing standard convolution into channel-wise convolution and pointwise convolution through depthwise separable convolution, the number of model parameters and computational complexity are significantly reduced, making it suitable for edge industrial control equipment with limited computing power.

[0040] Because the standard EfficientNet model defaults to a three-channel RGB image as input (input dimension... The input data for this invention is a single-channel grayscale image output by the multi-dimensional data optimization preprocessing module 10. (Input dimension) Therefore, the first convolutional layer structure of the backbone network needs to be adapted. The system changes the input channel parameter of the first convolutional layer from 3 to 1.

[0041] To fully utilize the weight parameters pre-trained on large-scale datasets (such as ImageNet), accelerate model convergence, and prevent gradient instability in the early stages of training, this embodiment employs a mean weight transfer initialization strategy. Specifically, let the weights of the first-layer convolutional kernel of the pre-trained model on the RGB channels be as follows: , , The newly constructed single-channel convolution kernel weights The calculation is as follows: .

[0042] The physical principle behind this initialization method is to maintain the amplitude distribution of the convolutional kernel's response to the input signal approximately consistent with that in the pre-training state, enabling the network to possess basic edge extraction capabilities from the early stages of training without having to learn from scratch. The adapted backbone network can directly process grayscale image tensors of size 640×640×1. The specific network depth, width scaling factor, and the specific layer configuration of the MBConv module in EfficientNet-B0 are well-known technical parameters in the field of deep learning, and those skilled in the art can obtain them from relevant open-source literature; therefore, they will not be elaborated upon here.

[0043] S202 generates a multi-scale feature pyramid output. The feature extraction backbone network module 20 extracts feature maps at three different levels through layer-by-layer downsampling operations, which are defined as shallow feature maps. Mid-layer feature map and deep feature maps These three feature levels constitute the feature pyramid, which not only decreases exponentially in spatial resolution but also increases non-linearly in the number of feature channels to encode information at different levels of abstraction. The specific definitions are as follows: Shallow feature map Corresponding to the third stage output of the backbone network, its downsampling factor is 8. The shallow feature map is [data missing] relative to the 640×640 input image. The spatial resolution is 80×80. This hierarchical feature map has high spatial fidelity and mainly encodes the edges, corners, and texture gradient information of objects. In the explosive body recognition task of this embodiment, The layer is responsible for responding to subtle geometric features such as detonator leads and printed text marks on the surface of explosive charges.

[0044] Mid-layer feature map Corresponding to the fourth stage output of the backbone network, its downsampling factor is 16. The mid-layer feature map, relative to the input image... The spatial resolution is 40×40. This hierarchical feature map is in an intermediate state between geometric features and semantic features, aggregating the combined information of local components, and is suitable for recognizing objects with specific shapes and structures, such as block emulsion explosives and the main body of a grenade.

[0045] Deep feature map Corresponding to the output of stage 5 of the backbone network, its downsampling factor is 32. The deep feature map, relative to the input image... The spatial resolution is 20×20. This hierarchical feature map has the largest receptive field and the highest semantic abstraction dimension, ignoring high-frequency texture details. It mainly represents the topological positional relationship between objects and the environment and the overall contour category, making it suitable for identifying boxed stacks of explosives or large improvised explosive devices.

[0046] The final output of the feature extraction backbone network module 20 is... The resulting feature set, which spans a multi-scale spatial range from 1 / 8 to 1 / 32, provides a hierarchical data foundation for gradient enhancement at specific levels in the GSAM module and cross-scale feature fusion in the BiFPN module in subsequent steps.

[0047] See attached document Figure 4 The gradient-guided structured attention module 30 implementation step S300 aims to leverage prior information about physical edges in image processing to enhance the neural network's ability to perceive geometric structures by introducing a hard inductive bias. This step, by embedding a gradient operator with fixed parameters, forces the network to focus on high-frequency edge information during the shallow feature extraction stage, thereby addressing the problem of overfitting to blurred textures in low-contrast downhole environments by purely data-driven models. This step specifically includes the following sub-steps: S301, perform differentiable gradient operator convolution operation. In this embodiment, the gradient-guided structured attention module 30 first receives shallow feature maps from the backbone network. To extract physically meaningful edge gradients, the gradient-guided structured attention module 30 pre-sets two non-learnable fixed convolutional kernels, corresponding to the horizontal kernels of the Sobel operator. With vertical direction core The system processes shallow feature maps. Each feature channel is subjected to a depthwise convolution operation.

[0048] The purpose of using depthwise convolution is to maintain the independence of each feature channel, avoid premature fusion of information between channels, and ensure that the geometric structure within each feature channel can be extracted independently. Unlike the weight parameters in conventional convolutional layers, which are continuously updated through backpropagation, the convolution kernel weights in this step are locked to the standard values ​​of the Sobel operator. (Horizontal convolution kernel) Convolution kernel in the vertical direction The definition matrix is ​​as follows: ; ; Through the above convolution operations, the system outputs horizontal gradient feature maps respectively. With vertical gradient feature map The calculation process is described as follows: ; ; in, This represents the shallow feature map of the input. This represents a depthwise convolution operation. This represents a feature map that includes horizontal edge information. This represents a feature map containing vertical edge information, and the output is... and Maintain and Same channel dimension.

[0049] S302, Generate the structural attention mask. After acquiring the horizontal and vertical gradient feature maps, the gradient-guided structured attention module 30 calculates the combined magnitude of the spatial gradients and maps it to the attention weight distribution. This addresses the issue that the traditional sigmoid function cannot output low weights under non-negative input (i.e., ... To address the problem that background noise cannot be effectively suppressed, this embodiment introduces learnable affine transformation parameters to renormalize the gradient magnitude.

[0050] The generation process of the attention mask follows the following computational logic: ; in, This indicates the generated structural attention mask in spatial coordinates. The weight value at the location; This represents the Sigmoid activation function, used to compress values ​​into the probability space (0,1). and Representing coordinates The horizontal and vertical gradient values ​​at the location. For learnable scaling factors, This is a learnable bias term (initialized to a negative value, e.g., -2.0, in this embodiment). By introducing... This shifts the gradient magnitude, which is close to 0 in the background region, to the negative interval before the input sigmoid function, thus affecting the output. Approaching 0, it effectively suppresses the smooth background; while the high gradient magnitude in the edge region is achieved through... Magnify, make A value close to 1 enables an attention mechanism for geometric edges.

[0051] S303, Perform feature residual enhancement operation. To apply the generated structural attention mask to the original features while avoiding feature value distribution collapse caused by multiple convolutions and mask multiplications, this embodiment uses residual connections for feature fusion. Compared with the original shallow feature map Performing element-wise multiplication is physically equivalent to spatially weighting the original features using the calculated edge probability map, thereby enhancing the feature response values ​​of the edge regions.

[0052] Subsequently, the weighted features are added to the original features to obtain the final enhanced shallow feature map. The specific feature enhancement calculation formula is as follows: ; in, This represents the shallow feature map after geometric enhancement. This represents a shallow feature map of the original input. This represents element-wise multiplication. This represents the structural attention mask generated in step S302.

[0053] The output feature map is obtained through the above residual enhancement operation. Not only does it preserve the original texture information, but it also explicitly enhances the geometric boundary features of objects. This allows subsequent detection networks to obtain stronger feature responses when processing fine structures such as detonator pins and creases in explosive charges, thereby effectively improving the recognition accuracy for small targets or targets with blurred edges.

[0054] The brightness invariant feature alignment module 40 implements step S400 to eliminate the feature distribution offset introduced by the random brightness adjustment in step S100 through normalization, ensuring that the feature map input to the subsequent fusion network has uniform statistical distribution characteristics. This step decouples the image's content features from its illumination distribution characteristics by introducing an instance normalization mechanism before feature fusion, and specifically includes the following sub-steps: S401, Construct a feature alignment layer based on instance normalization. In this embodiment, the brightness invariant feature alignment module 40 is configured between the feature extraction backbone network and the bidirectional feature pyramid network (BiFPN). The system performs feature alignment on each level of the feature map output by the backbone network (i.e., Each instance normalization layer is constructed independently to maintain the spatial semantic independence of features at each scale.

[0055] The technical principle behind instance normalization lies in the fact that the random brightness enhancement strategy in step S100 causes significant differences in illumination statistics among different image samples within the same training batch (mean and variance distributions are inconsistent). If conventional batch normalization is used, its mechanism of calculating global statistics based on the entire batch of samples will be affected by the distribution jitter imposed during this preprocessing stage, resulting in the normalized features failing to accurately reflect the inherent geometric structure of a single image. Instance normalization, on the other hand, standardizes only the spatial statistics of the current single image, effectively removing the specific illumination bias of that image and achieving adaptive correction of the feature distribution.

[0056] S402, calculates the channel statistics for the spatial dimension. For any input feature map tensor... (e.g. from) The hierarchical features), whose dimensions are defined as follows: ,in Indicates batch size, Indicates the number of channels. and These represent the height and width of the feature map, respectively. The brightness-invariant feature alignment module 40 is for the first... The first sample Each channel, in the spatial dimension Calculate the mean With variance .

[0057] mean The calculation formula is as follows: ; variance The calculation formula is as follows: ; in, Indicates the input feature map at the th The sample, the first Each channel, spatial coordinates The above calculation process ensures that the statistical value is obtained solely based on the pixel distribution of the current image, thus eliminating information interference paths between different samples within a batch.

[0058] S403, Perform feature re-standardization and affine transformation. After obtaining the independent statistics for each channel, the brightness invariant feature alignment module 40 first performs zero-mean unit variance standardization on the features, and then introduces learnable affine transformation parameters to restore the feature representation ability that is beneficial to the recognition task.

[0059] The calculation formulas for feature standardization and reconstruction are as follows: ; in, This represents the feature value output after alignment. This is the numerical stability constant, which is taken as 1×10 in this embodiment. -5 Used to prevent variance Calculation overflow caused by the invalidation of the denominator when it approaches 0; and The first Each channel corresponds to a learnable scaling factor and a translation factor. During this process, the normalization operation... It effectively filters out global brightness and contrast shifts caused by changes in external lighting. Parameters and These parameters are set to 1 and 0 respectively during network initialization and are updated via backpropagation during training. The physical significance of these two parameters is that they allow the network to autonomously learn and recover the texture contrast information crucial for explosive body identification based on normalized features, thereby avoiding the loss of semantic features caused by forced normalization. The feature map processed through steps S401 to S403 possesses invariance to changes in input illumination, providing a consistent data foundation for subsequent cross-scale feature fusion.

[0060] See attached document Figure 5 The implementation step S500 of the bidirectional feature pyramid fusion module 50 aims to address the semantic bias problem caused by directly adding features of different scales with equal weights in traditional feature pyramids (FPN). Since deep features emphasize semantic categories and shallow features emphasize geometric contours, simple linear superposition often obscures crucial, minute target signals. This step constructs bidirectional cross-scale connections through the bidirectional feature pyramid fusion module 50 and introduces a learnable, fast normalization fusion mechanism, transforming the brightness-aligned feature map output from step S400... Efficient aggregation is performed to generate fused features that combine high-resolution geometric details with strong semantic contextual information. This step specifically includes the following sub-steps: S501, perform channel dimension unification and spatial scale alignment. Since the feature maps output by the backbone network at different levels typically have inconsistent channel numbers and varying spatial resolutions, direct fusion would lead to dimensionality mismatch. In this embodiment, the bidirectional feature pyramid fusion module 50 first performs pointwise convolution on the input feature maps at each level using a 1×1 convolution kernel. This operation not only unifies the number of channels in all feature layers participating in the fusion to the preset dimension, but also... (In this embodiment, in order to balance accuracy and inference speed on the downhole embedded device, the following settings are made) It also serves as a linear projection, mapping features from different levels to the same feature vector space.

[0061] To address the resolution conversion involved in feature fusion, the bidirectional feature pyramid fusion module 50 employs a differentiated resampling strategy: for top-down paths, i.e., from low-resolution high-level features (such as... ) to high-resolution low-level features (such as During the transmission, nearest neighbor interpolation or bilinear interpolation is used for upsampling; for bottom-up paths, i.e., from high-resolution low-level features (such as... ) to high-level features at low resolution (such as During transmission, a max pooling layer is used for downsampling.

[0062] S502, Construct a redundant bidirectional fusion topology. To improve feature propagation efficiency, this embodiment optimizes the traditional unidirectional structure of FPN. The optimization strategy involves removing intermediate nodes with only a single input edge. The underlying principle is that single-input nodes do not perform multi-source feature aggregation, only increasing computational latency without contributing substantial feature enhancement. Based on this, the bidirectional feature pyramid fusion module 50 constructs two parallel feature propagation paths: one is a top-down semantic enhancement path, which propagates deep features... The information was gradually transmitted to and One approach is to imbue shallow features with semantic category information; another is a bottom-up localization enhancement path, which enhances shallow features. Edge texture information is gradually passed to and This supplements the spatial location information lost in deep features.

[0063] Furthermore, to avoid the dissipation of original feature information during multi-layer convolution, the bidirectional feature pyramid fusion module 50 establishes direct skip connections between input and output nodes at the same level. For example, for At the hierarchical level, the fusion node not only receives from... upsampling features and from The downsampling features also directly receive the original data. Input features. This peer-to-peer connection mechanism physically constructs a direct information path, ensuring that the inherent feature representation of this level is not overly smoothed or diluted in complex cross-level fusion calculations.

[0064] S503 performs fast normalized weighted fusion. Given that feature maps of different resolutions contribute differently to the final object recognition (e.g., ...), ... It made significant contributions to small targets such as detonators, while (While it makes a significant contribution to large targets such as boxed explosives), simple addition is not the optimal solution. The bidirectional feature pyramid fusion module 50 introduces a fast normalization fusion algorithm based on learnable weights.

[0065] by Taking a hierarchical fusion node as an example, suppose it receives three input features: the original input from the same level after channel adjustment. From high-level ( Input after upsampling and from lower levels ( Input after downsampling Then the output fused features are... The calculation formula is as follows: ; in, This represents the fused output feature map; , , , where represents the corresponding learnable weight parameters, and , respectively, represent the importance contribution of features from different sources to the final result.

[0066] To ensure the physical meaning of the fusion weights (i.e., weights represent contribution ratios) and maintain numerical stability, this embodiment re-evaluates all weight parameters before each weighted calculation. Apply the ReLU activation function to enforce constraints. In the formula It is a very small numerical stability constant (taken in this embodiment). This is used to avoid calculation anomalies caused by a denominator of zero. Compared with the traditional Softmax-based normalization method, this formula abandons the computationally intensive exponential operation and only uses addition and division to achieve the weight normalization constraint. This significantly improves the deployment efficiency of the bidirectional feature pyramid fusion module 50 on FPGA or low-power ARM chips, meeting the timeliness requirements of real-time processing of downhole video streams.

[0067] The target detection output and regression calculation implementation step S600 aims to convert the multi-scale fused feature map output from step S500 into specific explosive target category information and precise spatial location coordinates. This step is executed by the target detection output module 60, which uses a convolutional sub-network with shared weights to perform dense prediction on the feature map and uses a decoding algorithm to restore the relative offsets to absolute coordinates. Specifically, it includes the following sub-steps: S601, Construct a decoupled detection head network structure. In this embodiment, the target detection output module 60 targets each level of feature map output by the bidirectional feature pyramid network. The exact same prediction logic is applied to each. To resolve the feature conflict between the two types of tasks in object detection, namely, the classification task requires features to be translation invariant, so that the object should be identified as the same category regardless of its location; while the localization task requires features to be translationally equivariant, so that the movement of the object's position should cause a corresponding change in the regression coordinates, the object detection output module 60 is designed as two parallel sub-networks with independent parameters and decoupled structure: a classification sub-network and a box regression sub-network.

[0068] Both the classification subnetwork and the box regression subnetwork consist of multiple stacked depthwise separable convolutional layers. In this embodiment, each subnetwork contains 3 to 4 convolutional layers with a kernel size of 3×3 and the activation function being Swish or SiLU. This design significantly reduces the number of model parameters and computational complexity while ensuring that the receptive field covers the target region. Furthermore, different feature pyramid levels ( to The weight parameters of the same target detection output module 60 are shared. This weight sharing mechanism forces the network to learn a general feature representation across scales, enabling the model to detect both large and small explosive targets using the same set of parameters.

[0069] S602 generates the category confidence prediction result. After receiving the feature map, the classification sub-network of the object detection output module 60 outputs a dimension of... The tensor of . Among them, and The height and width of the current layer feature map; The preset number of detection categories (in the downhole scenario of this embodiment, (These correspond to emulsion explosives, industrial detonators, and explosive packaging boxes, respectively). The number of anchor frames preset for each spatial location (e.g.) (Preset frames corresponding to different aspect ratios and sizes).

[0070] The object detection output module 60 applies a Sigmoid activation function to the last dimension of the output tensor, mapping the logical values ​​of the convolution output to probability values ​​in the (0,1) interval. For positions on the feature map... and the corresponding first Anchor boxes, category scores This represents the posterior probability that a target of a specific category exists within the anchor frame.

[0071] S603 performs bounding box regression and coordinate decoding. The output dimension of the box regression sub-network of the object detection output module 60 is... The tensor represents the predicted bounding box's coordinates, but not its actual coordinates. Instead, it represents the offset of the predicted bounding box relative to a preset anchor box. To obtain the final absolute coordinates on the input image, the object detection output module 60 needs to perform a regression calculation operation.

[0072] Before performing the calculation, the object detection output module 60 first determines the feature map grid points based on the downsampling ratio (i.e., step size) of the current feature map layer relative to the original image. Map back to the center coordinates of the original image. Assume the center coordinates of the preset anchor frame are... Width is The height is The four offsets output by the network regression branch are respectively The final predicted bounding box center coordinates are... and its width and height Follow the regression solution formula as follows: ; ; ; ; in, This represents the x and y coordinates of the center point of the predicted bounding box in the image coordinate system after calculation. and This represents the width and height of the calculated prediction box; and This represents the translation scaling factor predicted by the network relative to the anchor frame size; and This represents the logarithmic scaling factor for the size of the network predictions; It is a natural constant. Using the exponential function... and The calculation logic is to ensure that the predicted width and height are always positive, so as to avoid generating negative-sized bounding boxes that do not exist physically.

[0073] S604, perform Non-Maximum Suppression (NMS) post-processing. Since the network predicts multiple overlapping bounding boxes around the same target, the target detection output module 60 needs to filter redundant results using the NMS algorithm. The system first removes predicted boxes with classification confidence scores below a preset confidence threshold (in this embodiment, this threshold is empirically set to 0.05 to filter background noise). Then, it sorts the bounding boxes from highest to lowest confidence score and calculates the Intersection over Union (IoU) between the highest-scoring box and the remaining boxes. If the IoU value of two bounding boxes is greater than a preset overlap threshold (in this embodiment, this threshold is empirically set to 0.5, indicating that when the overlap area exceeds half, it is considered a duplicate detection), it is determined to be a redundant prediction of the same target, and the box with the lower confidence score is removed. This process is repeated until all candidate boxes have been processed, and the final output set of bounding boxes is the final detection result of step S600.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying explosive targets based on multi-dimensional data optimization and feature fusion, characterized in that, Includes the following steps: The original image to be identified is acquired, and HSV spatial brightness adjustment, random cropping and grayscale conversion are performed on the original image to generate a single-channel grayscale image. A convolutional neural network is used to perform multi-level feature extraction on the single-channel grayscale image, outputting shallow feature maps, mid-level feature maps, and deep feature maps; Edge information of the shallow feature map is extracted using a fixed gradient operator to generate a structural attention mask, and the structural attention mask is applied to the shallow feature map to obtain an enhanced shallow feature map. The spatial dimension statistics of the enhanced shallow feature map, the middle feature map, and the deep feature map are calculated and re-standardized to generate a multi-scale feature map with distribution alignment. The channel dimension of the distributed and aligned multi-scale feature maps is unified, a bidirectional cross-scale transmission path is constructed and weighted fusion is performed to generate a fused multi-scale feature map. The decoupled prediction network is used to perform category prediction and bounding box regression based on the fused multi-scale feature map, and the explosive body detection results are output.

2. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 1, characterized in that, The step of performing HSV spatial brightness adjustment, random cropping, and grayscale conversion on the original image to generate a single-channel grayscale image includes: The original image is converted from the RGB color space to the HSV color space, keeping the hue and saturation components unchanged, and only applying a linear transformation to the lightness component; The adjusted brightness component is merged with the original hue and saturation components and then inversely transformed back to the RGB color space. Read the annotation file of the original image, randomly generate a cropping window under the constraint of satisfying a preset coverage ratio threshold, and scale the cropping area to a preset network input size; The weighted average method is used to calculate the weighted sum of the pixel values ​​of each RGB channel, and the color image is converted into the single-channel grayscale image.

3. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 1, characterized in that, The step of using a convolutional neural network to perform multi-level feature extraction on the single-channel grayscale image, and outputting shallow feature maps, mid-level feature maps, and deep feature maps includes: Construct a convolutional neural network that includes depthwise separable convolution and channel attention mechanisms, and adapt the input channel parameters of the first convolutional layer to a single channel; The weights of the first-layer convolutional kernel are initialized using the mean weight transfer strategy to keep the amplitude distribution of the convolutional kernel's response to the single-channel input signal approximately consistent with the pre-training state. Through layer-by-layer downsampling operations, the shallow feature map with a downsampling factor of 8, the middle feature map with a downsampling factor of 16, and the deep feature map with a downsampling factor of 32 are output respectively.

4. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 1, characterized in that, The step of extracting edge information from the shallow feature map using a fixed gradient operator to generate a structural attention mask, and applying the structural attention mask to the shallow feature map to obtain an enhanced shallow feature map, includes: Using preset, non-learnable horizontal and vertical convolutional kernels, depthwise convolution operations are performed on the shallow feature maps respectively, outputting horizontal gradient feature maps and vertical gradient feature maps; Calculate the combined gradient magnitude of the horizontal gradient feature map and the vertical gradient feature map at corresponding spatial locations; The gradient synthesis magnitude is renormalized by applying learnable affine transformation parameters and mapped to the structure attention mask with values ​​between 0 and 1 via a nonlinear activation function. The structural attention mask is multiplied element-wise with the shallow feature map, and the result is added to the original shallow feature map to obtain the enhanced shallow feature map.

5. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 1, characterized in that, The step of calculating the spatial dimension statistics of the enhanced shallow feature map, the middle feature map, and the deep feature map respectively, and then re-standardizing them to generate a distribution-aligned multi-scale feature map includes: An instance normalization layer is constructed independently for the feature map of each level; For each sample and each channel in the feature map, calculate the pixel mean and pixel variance of the channel in the spatial dimension. The feature pixel values ​​of the channel are standardized with zero mean and unit variance using the pixel mean and pixel variance. The standardized features are subjected to affine transformation using learnable scaling and translation factors to output the distribution-aligned multi-scale feature map.

6. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 1, characterized in that, The process of unifying the channel dimensions of the distributed, aligned multi-scale feature maps, constructing a bidirectional cross-scale transfer path, and performing weighted fusion to generate a fused multi-scale feature map includes: The input feature maps at each level are convolved point by point using 1×1 convolution kernels, and the number of channels in all feature maps is adjusted to the preset dimension. Remove intermediate nodes in the merged topology that have only a single input edge; Construct a top-down semantic enhancement path, and pass high-level features to low-level features after upsampling; Construct a bottom-up localization enhancement path, which downsamples low-level features and then passes them to high-level features; Establish direct jump connections between input and output nodes at the same resolution level; For each fusion node at each level, the input features from different paths are normalized and weighted by using learnable weight parameters.

7. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 1, characterized in that, The decoupled prediction network performs category prediction and bounding box regression based on the fused multi-scale feature map, and the output explosive particle detection results include: Construct a parameter-sharing classification subnetwork and a box regression subnetwork, both of which are composed of stacked depthwise separable convolutional layers; The classification subnetwork outputs logical values ​​indicating that each preset anchor box belongs to a specific category, and the activation function maps the logical values ​​to category confidence. The box regression subnetwork outputs the positional offset and size scaling factor of the predicted box relative to the preset anchor box; Based on the downsampling factor of the feature map and the geometric parameters of the preset anchor box, the absolute center coordinates and width and height dimensions of the prediction box in the original image are calculated using the position offset and the size scaling factor. Non-maximum suppression is applied to all predicted boxes based on the cross-union ratio threshold to remove redundant overlapping detection boxes.

8. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 2, characterized in that, The application of a linear transformation only to the lightness component includes: Generate random brightness adjustment coefficients that follow a uniform distribution; The original luminance component is linearly gained using the random luminance adjustment coefficient, and the calculation result is restricted to the minimum and maximum pixel values ​​allowed by the image format using a numerical truncation function. The random generation of the cropping window under the constraint of satisfying a preset coverage ratio threshold includes: Calculate the area of ​​the intersection between the randomly generated cropping window region and the target true bounding box region; Determine whether the ratio of the intersection area to the target true bounding box region is greater than or equal to a preset coverage ratio threshold. If the ratio is satisfied, retain the clipping window.

9. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 4, characterized in that, The step of performing depthwise convolution operations on the shallow feature map using preset, unlearnable horizontal and vertical convolution kernels includes: The weight matrix of the horizontal convolution kernel is set to the horizontal operator value of the Sobel operator, and the weight matrix of the vertical convolution kernel is set to the vertical operator value of the Sobel operator. The weights of the horizontal and vertical convolutional kernels are not updated during training. The step of renormalizing the gradient synthesis magnitude by applying learnable affine transformation parameters and mapping it via a nonlinear activation function to the structure attention mask with values ​​between 0 and 1 includes: Multiply the gradient composite magnitude by the learnable scaling factor and then add the learnable bias term; The weighted result is input into the Sigmoid function, so that the gradient magnitude of the background region is mapped to a range close to 0, and the gradient magnitude of the edge region is mapped to a range close to 1.

10. The explosive target recognition method based on multi-dimensional data optimization and feature fusion according to claim 6, characterized in that, The step of normalizing and weighting the input features from different paths using learnable weight parameters for each level of fusion node includes: For each input feature, the ReLU activation function is applied to the learnable weight parameter to force the weight parameter to be non-negative. Calculate the sum of the products of each input feature and its corresponding weight parameter; Calculate the sum of all the weight parameters and add the numerical stability constant to the sum; Divide the sum of the products by the sum after adding the numerical stability constant to obtain the fused feature output.

Images