Target real-time detection method and system fusing edge perception and interlayer feature enhancement

Abstract

The invention discloses a target real-time detection method and a target real-time detection system fusing edge perception and interlayer feature enhancement, relates to the technical field of computer vision and artificial intelligence, and discloses the target real-time detection method and the target real-time detection system fusing edge perception and interlayer feature enhancement. By combining a multi-scale feature fusion mechanism, the method can solve the problems of edge feature loss, insufficient inter-layer relevance and poor real-time performance in infrared small target detection, and has the advantages of improving the infrared small target detection precision and real-time performance, and remarkably reducing the false detection rate and omission ratio under complex background interference.

Description

Technical Field

[0001] This application relates to the fields of computer vision and artificial intelligence, and in particular to a real-time target detection method and system that integrates edge perception and inter-layer feature enhancement. Background Technology

[0002] With the rapid development of drones and infrared imaging technology, their applications in agricultural monitoring, security patrols, disaster early warning, and military reconnaissance are becoming increasingly widespread. Compared with visible light images, infrared images have advantages such as all-weather imaging and strong anti-interference capabilities, but they also face challenges such as low resolution, high noise interference, and small target size. In particular, small targets often occupy only a very small pixel area in the image, significantly increasing the difficulty of detection. Traditional image processing methods achieve separation by analyzing the characteristic models of the target and background, and some methods transform the detection problem into an optimization problem based on the sparsity of the target. Although these methods are easy to implement in engineering, they rely heavily on manually designed features and rules for judgment, making it difficult to adapt to complex environments and variable targets, and resulting in poor generalization ability.

[0003] In recent years, deep learning technology has made significant progress in the field of object detection, with multi-layer neural networks demonstrating powerful feature extraction capabilities. However, existing methods still have significant shortcomings in infrared small target detection tasks: on the one hand, traditional convolutional networks are prone to smoothing or even losing weak boundary signals during feature extraction, resulting in insufficient expression of small target edge features; on the other hand, existing feature fusion mechanisms are difficult to effectively capture the correlation between inter-layer features, limiting the model's ability to detect targets at multiple scales. Furthermore, mainstream detection algorithms struggle to balance computational complexity and inference speed, failing to meet the real-time requirements of edge devices. Especially under complex background interference, existing methods are prone to false positives and false negatives, severely affecting the reliability of detection results.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a real-time target detection method and system that integrates edge perception and inter-layer feature enhancement, aiming to improve the accuracy and real-time performance of infrared small target detection and significantly reduce the false detection rate and false negative rate under complex background interference.

[0006] To achieve the above objectives, this application proposes a real-time target detection method that integrates edge perception and inter-layer feature enhancement. The method includes:

[0007] Raw infrared image data is acquired using an infrared sensor;

[0008] Preprocessing operations are performed on the original infrared image data to obtain preprocessed image data; the preprocessing operations include at least one of rotation, random cropping, scaling, and normalization.

[0009] Based on the preprocessed image data, edge response map data is generated using an edge detection operator;

[0010] The preprocessed image data and the edge response map data are concatenated to obtain dual-channel input data;

[0011] The dual-channel input data is input into the backbone network, and the feature fusion module is used to perform feature enhancement processing on the context feature map data output by the backbone network to obtain enhanced feature map data.

[0012] Based on the enhanced feature map data, feature fusion and target detection processing are performed using a multi-scale detection head to obtain target detection result data.

[0013] In one embodiment, the step of generating edge response map data based on the preprocessed image data using an edge detection operator includes:

[0014] Based on the preprocessed image data, a horizontal Sobel convolution kernel is used to perform convolution operations on the data to obtain horizontal gradient data.

[0015] Based on the preprocessed image data, a vertical Sobel convolution kernel is used to perform convolution operations on the data to obtain vertical gradient data.

[0016] The horizontal gradient data and the vertical gradient data are squared respectively to obtain horizontal squared data and vertical squared data.

[0017] The edge response map data is obtained by adding the horizontal squared data and the vertical squared data and then taking the square root.

[0018] In one embodiment, the step of inputting the dual-channel input data into the backbone network and using the feature fusion module to perform feature enhancement processing on the context feature map data output by the backbone network to obtain enhanced feature map data includes:

[0019] The dual-channel input data is input into the backbone network, so that the feature fusion module receives the context feature map data output by the backbone network;

[0020] The context feature map data is subjected to dimensionality reduction and feature extraction to obtain local feature data;

[0021] The local feature data is decomposed and convolved, including horizontal convolution and vertical convolution, to obtain convolutional feature data.

[0022] The convolutional feature data is processed by an activation function to obtain channel attention weight data.

[0023] The channel attention weight data and the context feature map data are weighted to obtain weighted feature data;

[0024] The weighted feature data and the context feature map data are subjected to a residual connection operation to obtain the enhanced feature map data.

[0025] In one embodiment, the step of performing dimensionality reduction and feature extraction processing on the context feature map data to obtain local feature data includes:

[0026] The context feature map data is subjected to spatial dimension reduction processing through average pooling to obtain dimensionality-reduced feature data.

[0027] The dimensionality-reduced feature data is subjected to feature extraction operation through a 1×1 convolution kernel to obtain the local feature data.

[0028] In one embodiment, the step of performing decomposition convolution processing on the local feature data, including horizontal convolution and vertical convolution, to obtain convolutional feature data includes:

[0029] The local feature data is convolved horizontally using a 1×N convolution kernel, where N is the kernel size, to obtain the horizontal feature data.

[0030] The local feature data is convolved vertically using an N×1 convolution kernel to obtain vertical feature data.

[0031] The horizontal feature data and the vertical feature data are combined into the convolutional feature data.

[0032] In one embodiment, the step of processing the convolutional feature data with an activation function to obtain channel attention weight data includes:

[0033] The vertical feature data is processed using a 1×1 convolution kernel to obtain intermediate feature data;

[0034] The intermediate feature data is non-linearly mapped using the Sigmoid activation function, with the output value limited to the range of 0 to 1, to obtain the channel attention weight data.

[0035] In one embodiment, the step of performing feature fusion and target detection processing based on the enhanced feature map data using a multi-scale detection head to obtain target detection result data includes:

[0036] Based on the enhanced feature map data, feature map data at multiple scales are generated through the backbone network, including a first resolution feature map, a second resolution feature map, a third resolution feature map, and a fourth resolution feature map;

[0037] The feature map data at the multiple scales are fused, including upsampling, downsampling and stitching operations, to obtain fused feature map data;

[0038] The fused feature map data is classified and bounding box regressed using a detection head to obtain the target detection result data.

[0039] In one embodiment, the step of fusing the feature map data at multiple scales includes upsampling, downsampling, and concatenation operations to obtain fused feature map data.

[0040] The fourth resolution feature map is upsampled to obtain upsampled feature data.

[0041] The first resolution feature map is downsampled to obtain downsampled feature data.

[0042] The upsampled feature data, the second resolution feature map, the third resolution feature map, and the downsampled feature data are concatenated to obtain the fused feature map data.

[0043] In one embodiment, the step of classifying and regressing the fused feature map data using a detection head to obtain the target detection result data includes:

[0044] The fused feature map data is convolved using a 3×3 convolution kernel.

[0045] The convolutional data is then normalized using a normalization layer.

[0046] The normalized data is then nonlinearly activated using the SiLU activation function.

[0047] The activated data is used for feature mapping through a 1×1 convolution kernel to generate decoupled classification probability data and bounding box coordinate data, so as to obtain the target detection result data.

[0048] Furthermore, to achieve the above objectives, this application also proposes a real-time target detection system that integrates edge perception and inter-layer feature enhancement. The system includes: a memory, a processor, and a real-time target detection program that integrates edge perception and inter-layer feature enhancement, stored in the memory and executable on the processor. The real-time target detection program that integrates edge perception and inter-layer feature enhancement is configured to implement the steps of the real-time target detection method that integrates edge perception and inter-layer feature enhancement.

[0049] This application provides a real-time target detection method and system that integrates edge perception and inter-layer feature enhancement. By integrating edge perception and inter-layer feature enhancement techniques and combining a multi-scale feature fusion mechanism, it can solve the problems of edge feature loss, insufficient inter-layer correlation and poor real-time performance in infrared small target detection. It has the advantages of improving the accuracy and real-time performance of infrared small target detection and significantly reducing the false detection rate and false negative rate under complex background interference. Attached Figure Description

[0050] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating an embodiment of the real-time target detection method that integrates edge sensing and inter-layer feature enhancement according to this application.

[0053] Figure 2 This is a schematic diagram of the structure of an embodiment of the real-time target detection system that integrates edge perception and inter-layer feature enhancement according to this application.

[0054] Explanation of icon numbers:

[0055] 10. Memory; 20. Processor.

[0056] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0057] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0058] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0059] In existing technologies, drones and infrared imaging technology are widely used in fields such as agricultural monitoring and security inspection. However, infrared images suffer from problems such as low resolution, high noise interference, and a small proportion of pixels for small targets. Traditional methods rely on manually designed features and rules for judgment, which are difficult to adapt to complex environments. Although deep learning models improve detection accuracy, they have drawbacks such as loss of edge information, high computational complexity, and insufficient inference efficiency. Especially in small target detection tasks, weak edge features are easily smoothed by convolution operations, resulting in blurred target contours and affecting detection performance.

[0060] To address the aforementioned issues, effectively preserving edge information and improving real-time performance are crucial for small target detection in infrared images. Traditional convolutional networks tend to lose edge details during feature extraction, while directly introducing an edge detection module increases computational burden. Analysis reveals that fusing edge response information with the original image and inputting it into the network strengthens boundary feature representation. Simultaneously, a lightweight feature fusion module is designed to enhance contextual information while reducing the number of parameters. Combined with a multi-scale detection mechanism, this approach balances detection accuracy and speed.

[0061] Based on this, embodiments of this application provide a real-time target detection method that integrates edge perception and inter-layer feature enhancement, referring to... Figure 1 The real-time target detection method that integrates edge perception and inter-layer feature enhancement includes steps S100 to S600, wherein:

[0062] Step S100: Acquire raw infrared image data using an infrared sensor;

[0063] Step S200: Perform preprocessing operations on the original infrared image data to obtain preprocessed image data; the preprocessing operations include at least one of rotation, random cropping, scaling, and normalization operations;

[0064] Step S300: Based on the preprocessed image data, generate edge response map data using an edge detection operator;

[0065] Step S400: The preprocessed image data and the edge response map data are spliced ​​together to obtain dual-channel input data;

[0066] Step S500: Input the dual-channel input data into the backbone network, and use the feature fusion module to perform feature enhancement processing on the context feature map data output by the backbone network to obtain enhanced feature map data;

[0067] Step S600: Based on the enhanced feature map data, feature fusion and target detection processing are performed using a multi-scale detection head to obtain target detection result data.

[0068] In this embodiment, preprocessing refers to geometric transformation and standardization of the original infrared image. Specifically, this can be achieved by random rotation within ±15 degrees and cropping ratios between 0.5 and 1.0, enhancing data diversity and standardizing input size. The edge detection operator extracts image edge features through directional gradient calculation. Specifically, the Sobel operator can be used to calculate horizontal and vertical gradients separately and then synthesize the edge response map, enhancing target contour information. Dual-channel input data involves concatenating the preprocessed image and the edge response map along the channel dimension. This can be achieved using matrix concatenation, allowing the network to simultaneously receive both original texture and edge features. The feature fusion module applies attention weighting and residual connections to the features output by the backbone network. This can be achieved through decomposition convolution and channel attention mechanisms, enhancing key features and suppressing redundant information. The multi-scale detection head fuses and predicts feature maps of different resolutions. This can be achieved through a combination of upsampling, downsampling, and feature concatenation, capturing the feature representations of targets of different sizes.

[0069] In this embodiment, the original infrared image undergoes preprocessing such as rotation and cropping. The gradient is then calculated using the Sobel operator to generate an edge response map, which is then stitched together to form a dual-channel input. When the contextual features extracted by the backbone network pass through the feature fusion module, local features are first extracted through spatial dimensionality reduction and 1×1 convolution. Then, directional features are captured through horizontal and vertical decomposition convolutions. Sigmoid activation generates channel attention weights, which are then used to weight the original features before outputting enhanced features via residual connections. The multi-scale detection head assembles and stitches feature maps from different levels after upsampling and downsampling to align resolution. Finally, the classification result and localization coordinates are decoupled and output through convolutional layers.

[0070] In this embodiment, a dual-channel input mechanism is used to input edge features and the original image into the network in parallel, enhancing boundary information from the early stages of feature extraction. The feature fusion module uses decomposed convolution to reduce the number of parameters and dynamically adjusts feature weights using a channel attention mechanism, enhancing key features while maintaining computational efficiency. The multi-scale feature fusion method avoids the loss of details caused by repeated downsampling, making it more suitable for small target detection scenarios. Thus, this application effectively improves the detection accuracy of small targets in infrared images, enhances target contour features through edge response maps, and alleviates the boundary blurring problem caused by low resolution. The feature fusion module enhances the correlation between inter-layer features while reducing computational resource consumption, meeting the real-time processing needs of edge devices. The multi-scale detection mechanism takes into account the feature representation of targets of different sizes and can maintain stable detection performance even under complex background interference.

[0071] In one feasible implementation, the step of generating edge response map data based on the preprocessed image data using an edge detection operator includes: performing a convolution operation on the data using a horizontal Sobel convolution kernel based on the preprocessed image data to obtain horizontal gradient data; performing a convolution operation on the data using a vertical Sobel convolution kernel based on the preprocessed image data to obtain vertical gradient data; squaring the horizontal gradient data and the vertical gradient data respectively to obtain horizontal squared data and vertical squared data; and adding the horizontal squared data and the vertical squared data and then taking the square root to obtain the edge response map data.

[0072] In this embodiment, the horizontal Sobel convolution kernel is used to detect horizontal edges of the image. It can be implemented using a 3×3 matrix template, with the central element having a weight of 0 and the left and right elements having symmetrical positive and negative weights. The vertical Sobel convolution kernel is used to detect vertical edges of the image. It can also be implemented using a 3×3 matrix template, with the central element having a weight of 0 and the top and bottom elements having symmetrical positive and negative weights. Square calculation refers to performing a numerical square operation on the gradient data, which can be implemented through pixel-by-pixel multiplication to eliminate the influence of gradient direction on the magnitude. Square root calculation refers to performing a square root operation on the sum of squares, which can be implemented using numerical calculation library functions to restore the actual magnitude of the gradient magnitude.

[0073] In this embodiment, in the infrared image small target detection scenario, because the target edge information is weak and easily interfered with by noise, it is necessary to enhance the edge features through multi-directional gradient calculation. First, Sobel convolution kernels in the horizontal and vertical directions are used to perform convolution operations on the preprocessed image to extract gradient components in two orthogonal dimensions. Then, the gradient data is squared to eliminate the influence of the difference between positive and negative gradient directions on the amplitude calculation, retaining the purely quantized gradient intensity information. Finally, the squared gradient data in the two directions are superimposed and the square root operation is performed to obtain a response map that integrates horizontal and vertical edge information. This process, through orthogonal decomposition and vector synthesis, completely preserves the geometric features of the target edge and avoids the edge breakage problem caused by single-directional gradient detection.

[0074] In this embodiment, multi-dimensional modeling of target edge features is achieved through a three-step process: orthogonal convolution kernel decomposition, gradient squaring to eliminate directional bias, and vector synthesis to restore amplitude. For example, the existing method of directly adding the absolute values ​​of horizontal and vertical gradients can lead to an overestimation of diagonal edge responses. However, this solution uses square root calculation, which better conforms to the vector superposition rules of image edges, effectively improving the detection accuracy of weak edges in infrared images. Especially in low-contrast scenes, the orthogonal gradient decomposition and vector synthesis mechanism avoids edge misjudgment caused by noise interference. At the same time, the square root calculation eliminates the influence of gradient direction on amplitude, making the edge response map more realistically reflect the geometric features of the target contour, providing more robust feature input for subsequent detection modules.

[0075] In one feasible implementation, the step of inputting the dual-channel input data into the backbone network and using the feature fusion module to perform feature enhancement processing on the context feature map data output by the backbone network to obtain enhanced feature map data includes: inputting the dual-channel input data into the backbone network, so that the feature fusion module receives the context feature map data output by the backbone network; performing dimensionality reduction and feature extraction processing on the context feature map data to obtain local feature data; performing decomposition convolution processing on the local feature data, including horizontal convolution and vertical convolution, to obtain convolutional feature data; performing activation function processing on the convolutional feature data to obtain channel attention weight data; performing a weighted operation on the channel attention weight data and the context feature map data to obtain weighted feature data; and performing a residual connection operation on the weighted feature data and the context feature map data to obtain the enhanced feature map data.

[0076] In this embodiment, dimensionality reduction and feature extraction refer to operations that reduce the spatial dimension of the feature map and extract key local features. Specifically, this can be achieved using average pooling combined with a 1×1 convolutional kernel. Average pooling compresses spatial information while preserving channel features, and the 1×1 convolutional kernel enables cross-channel feature interaction. Convolution decomposition refers to breaking down the standard convolution into independent convolution operations in both horizontal and vertical directions. Specifically, 1×N and N×1 convolutional kernels can be used to extract horizontal and vertical features respectively. Separating the convolutional kernel parameters reduces computation while enhancing the ability to capture directional features. Channel attention weight data refers to weight coefficients representing the importance of different channel features. Specifically, this can be generated by non-linearly mapping intermediate feature data using the Sigmoid activation function, constraining the weight values ​​to the range of 0 to 1 to quantify the inter-channel dependencies. Residual connection operation refers to the operation of superimposing weighted features with the original context feature map. Specifically, this can be achieved by element-wise addition, preserving the original feature information while introducing attention-guided feature enhancement.

[0077] In this embodiment, the context feature map is first compressed using an average pooling layer to reduce computational complexity while preserving global information across channels. Subsequently, a 1×1 convolutional kernel is used for feature transformation to extract cross-channel correlated local features. In the decomposition convolution stage, a horizontal 1×N convolutional kernel and a vertical N×1 convolutional kernel are used to process the feature map separately, where N can be an odd number such as 3 or 5. Separating the convolutional kernel parameters enables direction-sensitive feature extraction, enhancing the expressive power of edge and texture features. After the convolutional features are compressed using a 1×1 convolutional kernel, they are input into a Sigmoid function to generate channel attention weights. These weights are multiplied channel-by-channel with the original context feature map to highlight important feature channels and suppress redundant information. Finally, residual connections are used to superimpose the weighted features with the original features, fusing high-level semantic features while preserving low-level details to form an enhanced feature map.

[0078] In this embodiment, by combining decomposed convolution with a channel attention mechanism, the computational complexity is reduced while the representation of edge and direction-sensitive features is enhanced. The residual connection structure effectively avoids the attenuation problem of feature information during multi-layer transmission, making the enhanced feature map more conducive to capturing weak edge information in small target detection tasks. Thus, this application can effectively improve the edge feature representation capability of small targets in infrared images and alleviate the smoothing effect of traditional convolutional networks on weak edge signals during feature extraction. The channel attention mechanism enhances the saliency of target-related features by dynamically adjusting the feature channel weights, while the decomposed convolution operation strengthens the representation capability of edge structures through directional feature extraction. The residual connection ensures the integrity of low-level detail information during feature enhancement, thereby achieving more accurate small target localization and recognition under complex background interference.

[0079] In one feasible implementation, the step of performing dimensionality reduction and feature extraction processing on the context feature map data to obtain local feature data includes: performing spatial dimension reduction processing on the context feature map data through average pooling to obtain dimensionality-reduced feature data; and performing feature extraction operation on the dimensionality-reduced feature data through a 1×1 convolution kernel to obtain the local feature data.

[0080] In this embodiment, average pooling refers to the operation of compressing the spatial dimension of the feature map. Specifically, it can be implemented by averaging the pixel values ​​within the feature map region using a sliding window. This operation can reduce the resolution of the feature map, reduce subsequent computation, and retain key information about the spatial dimension. A 1×1 convolution kernel refers to a convolution operation with a kernel size of 1×1, which can be achieved by adjusting the mapping relationship between the number of input and output channels. This operation can extract the correlation features between channels, enhance the expressive power of local regions, and avoid introducing redundant parameters.

[0081] In this embodiment, in the feature fusion module, the context feature map is first input to an average pooling layer. The average pooling layer reduces the spatial size of the feature map by a preset ratio, for example, compressing the original feature map of size H×W to H / 4×W / 4 dimensionality-reduced feature data. This process significantly reduces the resolution of the feature map while preserving the main spatial information. Subsequently, the dimensionality-reduced feature data is input to a 1×1 convolutional layer. By adjusting the number of channels in the convolution kernel, for example, reducing the number of input channels from C to C / 2, feature channels are filtered and reorganized. After the 1×1 convolution operation, the output local feature data retains the key edge information in the original features while reducing noise interference, providing a more effective input for subsequent decomposition convolution processing.

[0082] In this embodiment, by cascading average pooling and 1×1 convolution, the spatial dimension is reduced while the feature correlation between channels is strengthened. This allows subsequent processing to focus on key information in the target region, avoiding redundant calculations caused by excessively high resolution and thus reducing computational resource consumption. This scheme enables the feature fusion module to maintain high detection accuracy and meet real-time processing requirements when running on edge devices, making it particularly suitable for infrared small target detection scenarios carried by UAVs.

[0083] In one feasible implementation, the step of performing decomposition convolution processing on the local feature data, including horizontal convolution and vertical convolution, to obtain convolutional feature data includes: performing horizontal convolution operation on the local feature data using a 1×N convolution kernel, where N is the kernel size, to obtain horizontal feature data; performing vertical convolution operation on the local feature data using an N×1 convolution kernel to obtain vertical feature data; and combining the horizontal feature data and the vertical feature data to form the convolutional feature data.

[0084] In this embodiment, deconvolution processing refers to splitting the standard two-dimensional convolution operation into two independent one-dimensional convolution operations: horizontal and vertical. Specifically, this can be achieved using 1×N and N×1 convolution kernels sliding along different directions. Separating the convolution directions reduces the number of parameters while preserving spatial dimensional information. Horizontal convolution operation involves feature extraction along the image width direction, for example, using a 1×3 or 1×5 convolution kernel to capture horizontal edges or texture structures. Vertical convolution operation involves feature extraction along the image height direction, for example, using a 3×1 or 5×1 convolution kernel to capture vertical edges or texture structures. The combination operation involves concatenating or element-wise adding the features extracted in the horizontal and vertical directions according to their channel dimensions, thereby forming a fused feature containing bidirectional detail information.

[0085] In this embodiment, in the infrared small target detection scenario, since the target edge information is weak and easily affected by background interference, decomposition convolution processing can enhance the edge response in the horizontal and vertical directions respectively. For example, when local feature data is input into the decomposition convolution module, the 1×N convolution kernel slides along the width direction to respond to changes in horizontal edge gradients, while the N×1 convolution kernel slides along the height direction to respond to changes in vertical edge gradients. After combining the two convolution results, the sensitivity of the feature map to the target contour can be enhanced without significantly increasing the computational load. Furthermore, this decomposition method can avoid the excessive smoothing of edge information by traditional two-dimensional convolution kernels during feature extraction, thereby more accurately locating the boundary of small targets.

[0086] In this embodiment, through the above technical solution, this application can significantly improve the edge feature representation capability of small targets in infrared images and enhance the model's ability to identify blurred contours. Under complex background interference, the bidirectional features extracted by decomposing convolution can effectively suppress noise response while preserving the key structural information of the target. This processing method reduces computational complexity, making the detection algorithm more suitable for the real-time deployment needs of edge devices, and has practical application value, especially for the detection of small targets in low-resolution infrared images.

[0087] In one feasible implementation, the step of processing the convolutional feature data with an activation function to obtain channel attention weight data includes: processing the vertical feature data through a 1×1 convolution kernel to obtain intermediate feature data; and performing non-linear mapping on the intermediate feature data through a Sigmoid activation function, with the output value limited to the range of 0 to 1, to obtain the channel attention weight data.

[0088] In this embodiment, a 1×1 convolution kernel refers to a convolution operation with a kernel size of 1×1. Specifically, it can be implemented using a convolutional layer with learnable parameters. Its function is to perform a linear transformation of the channel dimensions of the input features, enabling information exchange and dimensional adjustment between feature channels. The Sigmoid activation function is an S-shaped curve function, specifically implemented using the mathematical expression 1 / (1+e^{-x}). Its function is to map the input features to the interval between 0 and 1, generating channel attention weights to quantify the importance of different feature channels.

[0089] In this embodiment, during the feature enhancement process, the vertical feature data, after being processed by a 1×1 convolutional kernel, effectively integrates the correlation between channels, eliminates redundant information, and extracts key features. The intermediate feature data, after being non-linearly mapped using the Sigmoid activation function, generates an attention weight vector equal to the number of channels. Each element in this vector corresponds to a weight coefficient for a feature channel. The magnitude of the weight coefficient reflects the contribution of the corresponding channel to the object detection task; a larger weight coefficient indicates that the feature information of that channel is more important. By multiplying the channel attention weights with the original feature map channel by channel, the feature response of important channels can be adaptively enhanced, while suppressing interference from noise or irrelevant channels, thereby improving the model's ability to focus on key features.

[0090] In this embodiment, by combining decomposed convolution with the sigmoid activation function, dynamic feature calibration of the channel dimension can be achieved with low computational cost while preserving spatial details. Furthermore, the output range of the sigmoid function is limited to 0 to 1, making the adjustment process of attention weights more stable and avoiding gradient explosion or vanishing problems, which is beneficial for model training convergence. Thus, this application can effectively enhance the edge feature representation of small targets in infrared images, adaptively strengthening target-related feature channels through a channel attention mechanism and suppressing background noise interference. This scheme improves the detection accuracy of small targets while maintaining low computational complexity, making it suitable for real-time detection tasks on edge computing devices.

[0091] In one feasible implementation, the step of performing feature fusion and target detection processing based on the enhanced feature map data using a multi-scale detection head to obtain target detection result data includes: generating feature map data at multiple scales through the backbone network based on the enhanced feature map data, including a first resolution feature map, a second resolution feature map, a third resolution feature map, and a fourth resolution feature map; fusing the feature map data at multiple scales, including upsampling, downsampling, and concatenation operations, to obtain fused feature map data; and performing classification and bounding box regression processing on the fused feature map data using a detection head to obtain the target detection result data.

[0092] In this embodiment, multi-scale feature map data refers to feature maps at different resolution levels generated by the backbone network, such as first-resolution feature maps, second-resolution feature maps, third-resolution feature maps, and fourth-resolution feature maps. Specifically, this can be achieved using convolutional layers of different depths to capture feature information of targets of different sizes. Fusion processing refers to stitching together feature maps of different resolutions after upsampling and downsampling to the same scale. Specifically, bilinear interpolation can be used for upsampling, and max pooling for downsampling, enhancing feature representation capabilities through cross-scale information integration. The detection head refers to the network structure used for target classification and localization. Specifically, it can be implemented using a combination of convolutional layers, normalization layers, and activation functions, outputting target category probabilities and bounding box coordinates through decoupled classification and regression branches, respectively.

[0093] In this embodiment, the enhanced feature map data output by the backbone network is input to the multi-scale detection head, generating four feature maps with different resolutions. The fourth resolution feature map expands its spatial dimension through upsampling, while the first resolution feature map shrinks its spatial dimension through downsampling. The adjusted feature maps are then concatenated with the second and third resolution feature maps along the channel dimension to form fused feature map data. The fused feature map data undergoes feature extraction using a 3×3 convolutional kernel, the gradient distribution is stabilized by a normalization layer, and nonlinearity is introduced through the SiLU activation function. Finally, it is mapped by a 1×1 convolutional kernel to classify probability data and bounding box coordinate data, completing the output of the target detection result.

[0094] In this embodiment, multi-scale feature fusion effectively combines the detailed information of high-resolution feature maps with the semantic information of low-resolution feature maps, enhancing the model's sensitivity to small targets. Furthermore, existing detection methods rely on complex post-processing steps to remove redundant detection boxes, such as non-maximum suppression, while this scheme reduces redundant predictions and improves inference efficiency by integrating cross-scale information from fused feature maps. Simultaneously, the fused feature map directly outputs classification and localization results through decoupling the detection head, avoiding the computational overhead of traditional post-processing steps and meeting the real-time detection needs of edge devices.

[0095] In one feasible implementation, the step of fusing the feature map data at multiple scales includes upsampling, downsampling, and stitching operations to obtain fused feature map data. This step comprises: upsampling the fourth resolution feature map to obtain upsampled feature data; downsampling the first resolution feature map to obtain downsampled feature data; and stitching the upsampled feature data, the second resolution feature map, the third resolution feature map, and the downsampled feature data together to obtain the fused feature map data.

[0096] In this embodiment, upsampling refers to expanding the spatial dimension of low-resolution feature maps using interpolation algorithms or transposed convolutions. For example, bilinear interpolation is used to double the size of the feature map to recover high-resolution detail information. Downsampling refers to compressing the spatial dimension of high-resolution feature maps using pooling layers or convolutional layers. For example, max pooling is used to reduce the size of the feature map to half its original size to extract high-level semantic features. Concatenation refers to connecting feature maps of different scales along the channel dimension. For example, four feature maps of different resolutions are adjusted to the same size and then superimposed along the channel direction to achieve the integration and complementarity of multi-scale features.

[0097] In this embodiment, in the infrared small target detection task, the fourth-resolution feature map, after upsampling, can recover the detailed information of the target edge, compensating for the spatial information loss caused by multiple downsampling in deep networks. The first-resolution feature map, after downsampling, can enhance the expressive power of high-level semantic features and suppress background noise interference. By concatenating the four feature maps at different scales, the fused feature map simultaneously contains shallow detailed features, mid-level structural features, and high-level semantic features, enhancing the model's ability to perceive small targets. For example, the fourth-resolution feature map is aligned with the third-resolution feature map after bilinear upsampling, and the first-resolution feature map is aligned with the second-resolution feature map after downsampling through a convolutional layer with a stride of 2. Finally, the four feature maps are concatenated along the channel dimension to form a fused feature map containing multi-scale information.

[0098] In this embodiment, by simultaneously introducing upsampling and downsampling operations, a feature fusion path covering the entire resolution range is constructed, enabling bidirectional interaction between shallow high-resolution features and deep low-resolution features, effectively solving the detail loss problem caused by unidirectional information flow in traditional methods. Furthermore, through a cross-level feature stitching mechanism, the model can simultaneously capture target responses at different scales, significantly improving the detection stability of small targets in complex backgrounds, especially accurately distinguishing target edges from noise interference areas in low-resolution infrared images.

[0099] In one feasible implementation, the step of classifying and regressing the fused feature map data using a detection head to obtain the target detection result data includes: performing convolution operations on the fused feature map data using a 3×3 convolution kernel; normalizing the convolutional data using a normalization layer; performing non-linear activation on the normalized data using the SiLU activation function; and performing feature mapping on the activated data using a 1×1 convolution kernel to generate decoupled classification probability data and bounding box coordinate data to obtain the target detection result data.

[0100] In this embodiment, a 3×3 convolutional kernel refers to a two-dimensional filter with a kernel size of 3 rows and 3 columns. Specifically, it can be implemented using a convolution operation with a stride of 1 and padding of 1, used to extract local contextual information while maintaining the spatial resolution of the feature map. A normalization layer standardizes the feature map output by the convolution, specifically through batch normalization or layer normalization, used to alleviate internal covariate shift problems, accelerate model convergence, and improve training stability. The SiLU activation function is a non-linear function with self-gating properties, specifically implemented using linear units with Sigmoid weighted inputs, used to enhance feature representation and suppress gradient vanishing. A 1×1 convolutional kernel refers to a two-dimensional filter with a kernel size of 1 row and 1 column, specifically implemented through channel number adjustment and feature decoupling operations, used to map the fused high-dimensional features into two independent outputs: classification probability and bounding box coordinates.

[0101] In this embodiment, the fused feature map data is first processed by a 3×3 convolutional kernel. This captures the spatial correlation between adjacent pixels through local receptive fields, enhancing the distinction between target edges and background. Subsequently, a normalization layer standardizes the feature distribution of the convolutional output, ensuring that the data from each channel are within a similar numerical range and preventing saturation of the subsequent activation function output due to differences in feature scale. The normalized data is then input into the SiLU activation function, whose smooth non-linear characteristics help preserve weak edge signals while suppressing noise interference. Finally, a 1×1 convolutional kernel maps the activated high-dimensional features to classification and regression branches, respectively. The classification branch outputs the target class probability of each anchor box, while the regression branch outputs the center coordinates, width, and height offset of the bounding box, achieving decoupled output of the detection results.

[0102] In some specific implementations, the normalization layer can employ batch normalization, which standardizes the data based on the mean and variance of the current batch during the training phase and uses a moving average statistic during the inference phase. The SiLU activation function can be replaced with ReLU or LeakyReLU, but its advantage in gradient smoothness may be weakened. The number of output channels of the 1×1 convolutional kernel can be dynamically adjusted according to the number of target categories and the anchor box size. For example, in a 20-class target detection task, the classification branch can be set to 20 channels and the regression branch can be set to 4 channels.

[0103] In this embodiment, the combination of 3×3 convolution and 1×1 convolution reduces the number of parameters while ensuring feature extraction capability. The SiLU activation function is combined to enhance the response of small target edge features. By decoupling the output, the mutual interference between classification and regression tasks is reduced, thereby improving the classification accuracy of small targets in infrared images, reducing bounding box coordinate regression errors, and reducing computational complexity. This makes the detection model easier to deploy in embedded devices with limited computing power, meeting real-time processing requirements.

[0104] In this embodiment, the real-time target detection method that integrates edge perception and inter-layer feature enhancement, by combining edge perception and inter-layer feature enhancement techniques with a multi-scale feature fusion mechanism, can solve the problems of edge feature loss, insufficient inter-layer correlation and poor real-time performance in infrared small target detection. It has the advantages of improving the accuracy and real-time performance of infrared small target detection and significantly reducing the false detection rate and false negative rate under complex background interference.

[0105] A real-time target detection system integrating edge perception and inter-layer feature enhancement, reference Figure 2 The system includes: a memory 10, a processor 20, and a real-time target detection program that integrates edge perception and inter-layer feature enhancement, stored in the memory 10 and executable on the processor 20. The real-time target detection program that integrates edge perception and inter-layer feature enhancement is configured to implement the steps of the real-time target detection method that integrates edge perception and inter-layer feature enhancement.

[0106] The real-time target detection system integrating edge perception and inter-layer feature enhancement provided in this application, employing the real-time target detection method integrating edge perception and inter-layer feature enhancement described in the above embodiments, can improve the accuracy and real-time performance of infrared small target detection, and significantly reduce the false detection rate and false negative rate under complex background interference. Compared with the prior art, the beneficial effects of the real-time target detection system integrating edge perception and inter-layer feature enhancement provided in this application are the same as those of the real-time target detection method integrating edge perception and inter-layer feature enhancement provided in the above embodiments, and other technical features of the real-time target detection system integrating edge perception and inter-layer feature enhancement are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0107] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A real-time target detection method integrating edge perception and inter-layer feature enhancement, characterized in that, The method includes: Raw infrared image data is acquired using an infrared sensor; Preprocessing operations are performed on the original infrared image data to obtain preprocessed image data; the preprocessing operations include at least one of rotation, random cropping, scaling, and normalization. Based on the preprocessed image data, edge response map data is generated using an edge detection operator; The preprocessed image data and the edge response map data are concatenated to obtain dual-channel input data; The dual-channel input data is input into the backbone network, and the feature fusion module is used to perform feature enhancement processing on the context feature map data output by the backbone network to obtain enhanced feature map data. Based on the enhanced feature map data, feature fusion and target detection processing are performed using a multi-scale detection head to obtain target detection result data.

2. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 1, characterized in that, The step of generating edge response map data based on the preprocessed image data using an edge detection operator includes: Based on the preprocessed image data, a horizontal Sobel convolution kernel is used to perform convolution operations on the data to obtain horizontal gradient data. Based on the preprocessed image data, a vertical Sobel convolution kernel is used to perform convolution operations on the data to obtain vertical gradient data. The horizontal gradient data and the vertical gradient data are squared respectively to obtain horizontal squared data and vertical squared data. The edge response map data is obtained by adding the horizontal squared data and the vertical squared data and then taking the square root.

3. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 1, characterized in that, The step of inputting the dual-channel input data into the backbone network and using the feature fusion module to perform feature enhancement processing on the context feature map data output by the backbone network to obtain enhanced feature map data includes: The dual-channel input data is input into the backbone network, so that the feature fusion module receives the context feature map data output by the backbone network; The context feature map data is subjected to dimensionality reduction and feature extraction to obtain local feature data; The local feature data is decomposed and convolved, including horizontal convolution and vertical convolution, to obtain convolutional feature data. The convolutional feature data is processed by an activation function to obtain channel attention weight data. The channel attention weight data and the context feature map data are weighted to obtain weighted feature data; The weighted feature data and the context feature map data are subjected to a residual connection operation to obtain the enhanced feature map data.

4. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 3, characterized in that, The step of performing dimensionality reduction and feature extraction on the context feature map data to obtain local feature data includes: The context feature map data is subjected to spatial dimension reduction processing through average pooling to obtain dimensionality-reduced feature data. The dimensionality-reduced feature data is subjected to feature extraction operation through a 1×1 convolution kernel to obtain the local feature data.

5. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 3, characterized in that, The step of performing decomposition and convolution processing on the local feature data, including horizontal convolution and vertical convolution, to obtain convolutional feature data includes: The local feature data is convolved horizontally using a 1×N convolution kernel, where N is the kernel size, to obtain the horizontal feature data. The local feature data is convolved vertically using an N×1 convolution kernel to obtain vertical feature data. The horizontal feature data and the vertical feature data are combined into the convolutional feature data.

6. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 3, characterized in that, The step of processing the convolutional feature data with an activation function to obtain channel attention weight data includes: The vertical feature data is processed using a 1×1 convolution kernel to obtain intermediate feature data; The intermediate feature data is non-linearly mapped using the Sigmoid activation function, with the output value limited to the range of 0 to 1, to obtain the channel attention weight data.

7. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 1, characterized in that, The step of performing feature fusion and target detection processing based on the enhanced feature map data using a multi-scale detection head to obtain target detection result data includes: Based on the enhanced feature map data, feature map data at multiple scales are generated through the backbone network, including a first resolution feature map, a second resolution feature map, a third resolution feature map, and a fourth resolution feature map; The feature map data at the multiple scales are fused, including upsampling, downsampling and stitching operations, to obtain fused feature map data; The fused feature map data is classified and bounding box regressed using a detection head to obtain the target detection result data.

8. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 7, characterized in that, The step of fusing the feature map data at multiple scales, including upsampling, downsampling, and concatenation operations, to obtain fused feature map data includes: The fourth resolution feature map is upsampled to obtain upsampled feature data. The first resolution feature map is downsampled to obtain downsampled feature data. The upsampled feature data, the second resolution feature map, the third resolution feature map, and the downsampled feature data are concatenated to obtain the fused feature map data.

9. The real-time target detection method integrating edge perception and inter-layer feature enhancement as described in claim 7, characterized in that, The step of classifying and regressing the fused feature map data using a detection head to obtain the target detection result data includes: The fused feature map data is convolved using a 3×3 convolution kernel. The convolutional data is then normalized using a normalization layer. The normalized data is then nonlinearly activated using the SiLU activation function. The activated data is used for feature mapping through a 1×1 convolution kernel to generate decoupled classification probability data and bounding box coordinate data, so as to obtain the target detection result data.

10. A real-time target detection system integrating edge perception and inter-layer feature enhancement, characterized in that, The system includes: a memory, a processor, and a real-time target detection program that integrates edge perception and inter-layer feature enhancement, stored in the memory and executable on the processor, wherein the real-time target detection program that integrates edge perception and inter-layer feature enhancement is configured to implement the steps of the real-time target detection method that integrates edge perception and inter-layer feature enhancement as described in any one of claims 1 to 9.

Images