An infrared target detection method based on a bidirectional receptive field attention feature network

Abstract

The application discloses an infrared target detection method based on a bidirectional receptive field attention feature network, and realizes the method by constructing a model comprising a feature coding network, a feature fusion network and a detection head, so as to improve the detection precision and model generalization ability of small targets in an infrared remote sensing image. The feature coding network introduces a receptive field attention convolution module, independently weights different spatial positions in the receptive field, enhances fine-grained feature extraction, and solves the problem of insufficient multi-scale target feature extraction. Then, the feature coding network alternately stacks an RFAConv module and a standard convolution module, optimizes feature extraction and fusion operations, thereby maximally retaining and enhancing key features of the target, and solving the problem of poor adaptability to a complex background. The feature fusion network adopts a bidirectional feature pyramid network, bidirectionally and efficiently fuses high-level semantic information and low-level detail information, and solves the problem of information loss in the fusion process.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and computer vision technology, and particularly relates to an infrared target detection method based on a bidirectional receptive field attention feature network. Background Technology

[0002] With the continuous advancement of high-resolution remote sensing image technology, infrared remote sensing image analysis has become an important tool in various fields such as military target monitoring, geographic monitoring, and resource management. Utilizing deep learning techniques, especially convolutional neural networks (CNNs) and attention mechanisms, to automatically and accurately detect targets from these remote sensing images is currently a mainstream research direction in this field. To address the problem of large target scale variations in remote sensing images, researchers typically employ multi-scale detection methods such as Feature Pyramid Networks (FPNs).

[0003] However, existing infrared remote sensing image target detection technologies still face several unresolved problems and technical shortcomings in practical applications: First, the multi-scale feature extraction capability is insufficient. Most existing convolutional neural networks use convolutional kernels with fixed parameters, making them difficult to adapt to the characteristics of remote sensing images, such as huge differences in target scale and dense distribution of small targets. Although methods such as Feature Pyramid Network (FPN) have made some improvements, their performance still has bottlenecks in handling large-scale changes and small target detection.

[0004] Second, complex backgrounds and severe noise interference. Infrared remote sensing images often contain rich background information (such as terrain, vegetation, buildings, etc.) and strong high-frequency noise. When the target and the background have similar textures or infrared features, existing methods often struggle to distinguish them effectively, resulting in high false positive and false negative rates.

[0005] Third, the feature attention mechanism has limitations. Although some methods introduce attention mechanisms to enhance the accuracy of feature extraction, most attention mechanisms only perform simple feature weighting in the spatial or channel dimensions. This mechanism lacks independent and fine-grained attention to different locations in the feature space, limiting the network's ability to represent features of small targets and making it difficult to fully utilize the detailed information of the image.

[0006] Fourth, the feature fusion methods are simplistic and inefficient. Most current mainstream feature fusion strategies employ a top-down or bottom-up unidirectional fusion approach. This unidirectional information flow struggles to adequately balance the interaction between high-level semantic information and low-level spatial details, easily leading to the loss of crucial information during the fusion process. Furthermore, existing methods lack bidirectional interaction between multi-scale features, failing to effectively utilize the complementary information of features at different scales, thus impacting the final detection performance. Summary of the Invention

[0007] The purpose of this invention is to address the insufficient accuracy of current infrared remote sensing image target detection technology by proposing an infrared target detection method based on receptive field attention convolution and employing a bidirectional fusion feature network.

[0008] This invention addresses the issue of insufficient detection accuracy primarily through two approaches: First, it utilizes receptive field attention convolution, alternating it with traditional convolution and the C3 module, to deeply optimize the feature extraction process by increasing the receptive field and adding a weighting mechanism. Second, it employs a bidirectional fusion feature network, introducing an additional bottom-up path. This allows strong semantic information from higher levels to be passed down to guide lower-level features, while fine spatial details from lower levels can be supplemented upwards, achieving efficient bidirectional information flow across different scales. By fully utilizing the capabilities of the pre-trained model, it outputs accurate detection results and provides strong theoretical and practical support for solving problems such as insufficient multi-scale feature extraction, complex backgrounds and noise interference, insufficient feature attention, and limitations in feature fusion methods in infrared remote sensing image detection.

[0009] The objective of this invention is achieved through the following technical solution: an infrared target detection method based on a bidirectional receptive field attention feature network, the method comprising: A dataset was constructed by selecting multimodal remote sensing image data including infrared, synthetic aperture radar, and visible light, and then preprocessed. A feature encoding network is constructed by alternating receptive field attention convolutions, traditional convolutions, and C3 modules, with context aggregation performed at the network ends; data is input into the feature encoding network to obtain feature maps at different levels; The feature maps at different levels are passed through a feature fusion network, which includes a top-down path to pass high-level semantic information and a bottom-up path to supplement low-level spatial details. Finally, three optimized feature maps are generated and fed into three parallel detection heads, which make predictions on their respective feature maps. Loss calculation and backpropagation are performed based on the dataset to obtain the trained feature encoding network and feature fusion network. The trained model weights are loaded, and the obtained prediction results are decoded into image coordinates and processed into detection boxes.

[0010] Furthermore, the preprocessing includes: The image is preprocessed using standardization. Meanwhile, to address the noise and background texture interference present in the infrared image, wavelet transform is used to decompose the image at multiple scales, separating the low-frequency contour information and high-frequency details and noise, preserving the edge features of the target, and suppressing irrelevant high-frequency noise interference. Adaptive histogram equalization is used to improve the contrast between the target and the background in the infrared image. Data augmentation was performed on the dataset: geometric transformations were used to simulate the changes in target pose, position, and size caused by different viewpoints and distances in actual remote sensing images; the brightness, contrast, and saturation of the images were randomly adjusted; and image regions were randomly cropped at different scales to generate training samples containing targets of different sizes.

[0011] Furthermore, the feature encoding network specifically includes: a first standard convolutional layer, a first receptive field attention convolutional layer, a first C3 convolutional layer, a second receptive field attention convolutional layer, a second C3 convolutional layer, a third receptive field attention convolutional layer, a third C3 convolutional layer, a second standard convolutional layer, and a fourth C3 convolutional layer connected in sequence. SPPF is used at the end of the network. The features output by the fourth C3 convolutional layer are processed in parallel by multiple max pooling kernels of different sizes, and the results are concatenated. Finally, the second receptive field attention convolutional layer, the third receptive field attention convolutional layer, and the SPPF-processed result are used as high-resolution, medium-resolution, and low-resolution feature maps, respectively, and input into the feature fusion network.

[0012] Furthermore, the receptive field attention convolutional layer comprises three parallel sub-modules, specifically: Weight generation: The input feature map is first passed through a 3×3 average pooling layer, and then fed into a 1×1 grouped convolution. The output of this convolutional layer is shaped and passed through a softmax activation function to generate normalized attention weights for different spatial locations within the receptive field.

[0013] Feature generation: The input feature map is simultaneously fed into a 3×3 grouped convolution, and then subjected to batch normalization and ReLU activation function to generate the original spatial features; Weighting and Integration: The attention weights are multiplied element-wise with the spatial features to achieve dynamic weighting of the feature map, enabling the network to adaptively focus on more important regions. The weighted features are then passed through a standard 3×3 convolutional layer, batch normalization, and SiLU activation function to output the final enhanced features.

[0014] Furthermore, the top-down path specifically includes: The lowest resolution feature map output by SPPF is first passed through a receptive field attention convolutional layer for feature enhancement; The enhanced feature map is initially upsampled to match the medium-resolution feature map. The feature map after the initial upsampling is weighted and fused with the medium-resolution feature map to obtain the first fused feature; The first fusion feature is refined and enhanced using a C3 layer and a receptive field attention convolutional layer. The refined and enhanced features are then upsampled a second time to match their spatial dimensions with the high-resolution feature map. The feature map after secondary upsampling is weighted and fused with the high-resolution feature map to obtain the second fused feature; The second fused feature is processed using the C3 layer and then input into the first detection head used to detect small targets.

[0015] Furthermore, the bottom-up path specifically refers to: The secondary fusion features processed by the C3 layer are then processed by the receptive field attention convolutional layer and used as the starting point of the path. The initial fusion features of the path starting point features, after downsampling and refinement enhancement, are weighted and fused to obtain the third fusion feature with medium resolution; The third fusion feature is processed by the C3 layer and then input into the second detection head used to detect medium-sized targets. The third fusion feature processed by the C3 layer is processed using a receptive field attention convolutional layer; The processed features are downsampled and then weighted and fused with the low-resolution feature map to obtain the fourth fused feature, and the medium-resolution information is added back to the low-resolution path. The fourth fusion feature is processed using C3 and then input into the third detection head used for detecting large targets.

[0016] Furthermore, the three optimized feature maps are fed into three parallel detection heads, which perform predictions on their respective feature maps as follows: The three detection heads process the fused feature maps at different scales to adapt to targets of different sizes; each detection head is a lightweight Conv2d convolutional layer to generate the final prediction result, which includes bounding box coordinates, target confidence, and class probability.

[0017] Furthermore, the loss calculation and backpropagation are specifically as follows: An improved EIoU loss function is used to optimize the model's localization accuracy. Based on the CIoU loss function, the difference between width and height is calculated separately to replace the aspect ratio, directly minimizing the difference in width and height between the predicted target bounding box and the ground truth bounding box. Simultaneously, a loss function is introduced to address the imbalance between easy and difficult samples, thus achieving better localization results. The calculation formula is as follows: Where A and B represent the ground truth bounding box and the predicted bounding box, The intersection of A and B divided by the union of A and B. , Let A be the center point of B. , Let A be the width and height. , Let B be the width and height. , These are the width and height of the smallest bounding rectangle of A and B; Confidence loss Classification loss Both methods use binary cross-entropy loss, and the general formula for calculating binary cross-entropy loss is as follows: Where y is the true label, and for confidence loss Positive samples are represented by 1, and negative samples by 0; for classification loss 1 represents the target belonging to this class, and 0 represents it not belonging to this class. p is the probability predicted by the model. The total loss is the weighted sum of the EIoU location loss, confidence loss, and classification loss.

[0018] On the other hand, the specification also provides an infrared target detection device based on a bidirectional receptive field attention feature network, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the infrared target detection method based on a bidirectional receptive field attention feature network.

[0019] On the other hand, the specification also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the infrared target detection method based on a bidirectional receptive field attention feature network.

[0020] The beneficial effects of this invention are: 1. The Receptive Field Attention Convolution (RFAConv) module is introduced. By independently weighting different spatial locations within the receptive field, it overcomes the limitations of traditional convolution sharing parameters, solves the problems of insufficient multi-scale target feature extraction and insufficient feature attention, and significantly improves the feature recognition capability of small targets.

[0021] 2. By alternately stacking the RFAConv module and the C3 module in the feature encoding network, the feature extraction process is deeply optimized, achieving maximum preservation and enhancement of key features, and effectively solving the problems of poor adaptability and low robustness of the model under complex backgrounds and noise interference.

[0022] 3. A bidirectional feature pyramid network (BiFPN) was adopted. Through its bidirectional information paths from top to bottom and bottom to top and an adaptive weighted fusion strategy, it achieved efficient complementarity between high-level semantic information and low-level spatial details, solving the problem of information loss caused by traditional one-way fusion.

[0023] 4. An improved EIoU loss function was adopted, which improves the model's localization accuracy by directly minimizing the difference in width and height between the predicted bounding box and the true bounding box; combined with multimodal data training and advanced data preprocessing techniques, the model's generalization ability under different data distributions was effectively improved. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the internal structure of the Receive Field Attention (RFA) module provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the internal data flow of the BiFPN module provided in this embodiment of the invention; Figure 3 This is an overall structural block diagram of the infrared target detection method provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of an infrared target detection device provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] like Figure 3 As shown in the figure, the present invention provides an infrared target detection method based on a bidirectional receptive field attention feature network. The method includes four core steps: Step 1, data preparation and enhancement; Step 2, model architecture and forward propagation; Step 3, loss calculation and backpropagation; Step 4, model inference.

[0028] Step 1, Data Preparation and Augmentation, includes the following sub-steps: 1.1 To ensure the generalization ability of the model, multimodal remote sensing image data, including infrared, synthetic aperture radar (SAR) and visible light, are selected to construct the dataset. These datasets usually contain a large number of images with precise bounding boxes, which delineate the location and size of various military targets, geographic entities or resource objects. 1.2 Preprocess the dataset by normalizing the image size to a fixed size (e.g., 640x640) and mapping the pixel values ​​to the range [0.0, 1.0] to suit network computation; 1.3 The dataset is further processed by using wavelet transform to separate low-frequency contour information and high-frequency details and noise from the image, preserving the edge features of the target while suppressing irrelevant high-frequency noise interference; and adaptive histogram equalization (CLAHE) is used to improve the local contrast of the image and enhance the visibility of image details. 1.4 The dataset was expanded using data augmentation techniques, including simulating geometric transformations (rotation, flipping, scaling) at different shooting angles and distances, as well as simulating color perturbations under different lighting conditions. In particular, considering the dense concentration of small targets in remote sensing images, a multi-scale random cropping strategy was adopted to generate training samples containing targets of different sizes, forcing the model to learn to adapt to scale changes; Step two, model architecture and forward propagation, is the core computational process of the model, such as... Figure 3 As shown, the data flows sequentially through the feature encoding network, the bidirectional feature fusion network, and the detection head. This includes the following sub-steps: 2.1 The Feature Encoding Network (Backbone) extracts hierarchical deep features from the input image. It is a network composed of nine core modules stacked together, and innovatively alternates between Receptive Field Attention Convolution (RFAConv), traditional Convolution Conv, and the C3 module (CSPBottleneck) in YOLOv5. As data flows through the network, the spatial resolution is progressively downsampled, resulting in five levels of feature maps. At the end of the network, contextual information is aggregated to further enhance feature representation. The specific workflow is as follows: Initially (Conv), the input image tensor size is [B, 3, H, W] (B is the batch size, 3 is the number of channels); through traditional convolution Conv, the input image passes through a convolutional layer with a stride of 2, outputting a feature map P1, whose size is downsampled to [B, C1, H / 2, W / 2]. This layer is used to quickly reduce the spatial resolution and initially extract low-level features; Step 1 (RFAConv): Input the P1 feature map into the RFAConv module, such as... Figure 1As shown, the feature map first passes through a 3×3 average pooling layer, then a 1×1 grouped convolution (3 groups, no bias). The output of this convolutional layer is shaped and passed through a Softmax activation function to generate normalized attention weights W_att for different spatial locations within the receptive field. Simultaneously, the P1 feature map is fed into a 3×3 grouped convolution (3 groups), and then subjected to batch normalization and a ReLU activation function to generate the original spatial feature F_spatial. Finally, the attention weights W_att and the spatial feature F_spatial are multiplied element-wise to dynamically weight the feature map, allowing the network to adaptively focus on more important regions. The weighted feature map then passes through a standard 3×3 convolutional layer with a stride of 2, batch normalization, and a SiLU activation function to output the final feature map P2 with dimensions [B, C2, H / 4, W / 4]. In step 2 (C3), feature map P2 is refined by module C3, and is divided into two parts in the channel dimension. One part of the feature map goes through a series of "bottleneck layer" backbone branches, which are composed of 1×1 convolution, 3×3 convolution and residual connections; the other part of the feature map is directly retained without any operation. The feature map processed by the backbone branches and the directly retained feature map are concatenated in the channel dimension. Finally, the concatenated feature map goes through a convolutional layer to integrate information and output the final feature map. In step 3 (RFAConv), the feature map is downsampled again by the RFAConv module, and the output feature map P3 is generated with dimensions [B, C3, H / 8, W / 8]. P3 has a large spatial resolution and rich details, and is a key level for detecting small targets.

[0029] From step 4 to step 6, the C3 module and the RFAConv module are used alternately to continue feature extraction and downsampling; step 5 (RFAConv) outputs feature map P4 with dimensions [B, C4, H / 16, W / 16]; P4 has a larger receptive field and richer semantic information.

[0030] Step 7 (Conv) involves a final downsampling through a standard convolutional layer, outputting feature map P5 with dimensions [B, C5, H / 32, W / 32]. P5 has the largest receptive field and the strongest semantic information, primarily used for detecting large targets. Step 8 (C3) involves the final feature purification of P5; Step 9 (SPPF) introduces the SPPF (Spatial Pyramid Pooling - Fast) module at the end of the network. It processes P5 features in parallel using multiple max pooling kernels of different sizes and concatenates the results to capture multi-scale contextual information without increasing the amount of computation, thereby enhancing the model's adaptability to targets of different sizes.

[0031] 2.2 The multi-scale features (P3 / 8, P4 / 16, P5 / 32, etc.) output from steps 4, 6, and 9 of the backbone network are input into a feature fusion network (Neck) based on a Bidirectional Feature Pyramid Network (BiFPN). This network deeply integrates and optimizes the three different levels of feature maps output by the "feature encoding network". For example... Figure 2 As shown, the entire process transmits high-level semantic information through a top-down path, then supplements low-level spatial details through a bottom-up path, and finally generates three optimized feature maps which are then fed into the detection head; the specific workflow is as follows: Phase 1, top-down information fusion path, which passes down high-level, semantically strong features (such as P5 / 32) layer by layer and fuses them with low-level features to enhance the semantic expressive power of low-level features.

[0032] In step 10 (RFAConv), the lowest resolution feature map (P5 / 32) from step 9 (SPPF) is first enhanced by an RFAConv module, while the feature map dimension remains unchanged.

[0033] Step 11 (UpSample) upsamples the feature map output from step 10 so that its spatial size matches that of the medium-resolution P4 / 16 feature map.

[0034] Step 12 (BiFPN) is a key fusion node; it weights and fuses the upsampled features from step 11 with the P4 / 16 feature map from step 6 to generate a preliminary fused feature that combines high-level semantics and mid-level details.

[0035] Steps 13 to 14 (C3 False, RFAConv): Further refine and enhance the medium-resolution features fused in step 12, while keeping the feature map dimensions unchanged.

[0036] Step 15 (UpSample): Upsample the feature map output from step 14 again to match its spatial dimensions with the highest resolution P3 / 8 feature map.

[0037] Step 16 (BiFPN): This is the second key fusion node in the top-down path. It performs a weighted fusion of the upsampled features from Step 15 and the P3 / 8 feature map from Step 4.

[0038] Step 17 (C3 False): Perform final processing on the high-resolution features fused in Step 16. The resulting image will be sent directly to the first detection head (80x80) used to detect small targets.

[0039] Phase Two: Bottom-Up Information Supplementation Path. After completing the top-down path, the model immediately initiates a bottom-up path, passing low-level features containing rich spatial details and positioning information upwards layer by layer and fusing them with high-level features again to supplement the details of high-level features and enhance positioning capabilities.

[0040] Step 18 (RFAConv): The high-resolution features rich in semantic information output from Step 17 are processed through an RFAConv module and used as the starting point of the bottom-up path.

[0041] Step 19 (BiFPN): This is the first key fusion node in the bottom-up path. It performs a weighted fusion of the downsampled features processed in Step 18 with the mid-resolution features output from Step 14. This step supplements the mid-resolution path with high-resolution detail information.

[0042] Step 20 (C3 False): Perform final processing on the medium-resolution features fused in Step 19, and the result will be sent to a second detection head (40x40) for detecting medium-resolution targets.

[0043] Step 21 (RFAConv): The medium-resolution features output from step 20 are processed through an RFAConv module.

[0044] Step 22 (BiFPN): This is the second key fusion node in the bottom-up path. It performs a weighted fusion of the features processed in Step 21 (after downsampling) with the low-resolution features output from Step 10, supplementing the low-resolution path with mid-resolution information.

[0045] Step 23 (C3 False): Perform final processing on the low-resolution features fused in Step 22. The result will be sent to a third detection head (20x20) for detecting large targets.

[0046] Through this complete and meticulous bidirectional fusion process, the network ensures that the feature maps ultimately fed into the three detection heads fully integrate semantic and spatial information from three different levels: high, medium, and low, thereby greatly improving the detection accuracy of multi-scale targets.

[0047] 2.3 The fused feature map is fed into three parallel detection heads, which make predictions on their respective feature maps to effectively adapt to the detection needs of targets of different sizes. The 80x80 detection head is dedicated to detecting small targets. It uses a lightweight Conv2d convolutional layer to process the input feature map. This convolutional layer directly calculates on the 80x80 feature map to generate dense prediction results, which mainly include three parts: bounding box coordinates, target confidence, and class probability.

[0048] Step 3, Loss Calculation and Backpropagation: An improved EIoU (Efficient IoU) loss function is used, independently penalizing width and height differences. Specifically, the EIoU loss function is an improvement on the CIoU (Complete IoU) loss function. The CIoU loss function considers factors such as the distance between the center points of the predicted and ground truth boxes, aspect ratio, and overlapping area, and its formula is as follows: Where A and B represent the ground truth bounding box and the predicted bounding box, The intersection of A and B divided by the union of A and B. , Let A be the center point of B. This represents the Euclidean distance between the two center points. The length of the diagonal of the smallest bounding rectangle that can contain both A and B. Represents the weighting coefficient. Used to measure the consistency of the aspect ratio between A and B. , Let A be the width and height. , Let B be the width and height.

[0049] EIoU replaces aspect ratio by calculating the difference between width and height separately, directly minimizing the difference in width and height between the predicted target bounding box and the ground truth bounding box. It also introduces a loss function to address the imbalance between easy and difficult samples, thus achieving better localization results. Its calculation formula is as follows: in, , It is the width and height of the smallest bounding rectangle of A and B.

[0050] Confidence loss Classification loss Both methods use binary cross-entropy loss, and the general formula for calculating binary cross-entropy loss is as follows: Where y is the true label, and for confidence loss Positive samples are represented by 1, and negative samples by 0; for classification loss 1 represents the target belonging to this class, and 0 represents it not belonging to this class. p is the probability predicted by the model.

[0051] The total loss is a weighted sum of the EIoU localization loss, confidence loss, and classification loss. The total loss function is as follows: in =0.05, =1.0, =0.5; After calculating the total loss, all parameters in the network are updated through the backpropagation algorithm. This process is repeated until the model converges and the model weights are obtained.

[0052] Step four, model inference, includes the following sub-steps: 4.1 Model Loading and Preprocessing: First, load the trained model weights and set the model to inference mode. The images to be detected need to undergo the same preprocessing steps as during training (size normalization, numerical normalization, etc.) to ensure the consistency of data distribution and avoid performance degradation due to data format mismatch.

[0053] 4.2 The preprocessed image undergoes a complete forward propagation to generate the original, dense prediction results; then, after post-processing, the predicted relative coordinates are decoded into absolute coordinates on the image; the non-maximum suppression (NMS) algorithm is used to remove redundant and overlapping detection boxes for the same target, retaining only the one with the highest confidence; the confidence threshold of NMS is set to 0.5, and the IoU threshold can be set to 0.45.

[0054] 4.3 Output Results: The coordinates of the detection boxes after NMS filtering are restored to the original image size. The final output includes the detection results with category labels, confidence scores, and precise locations, which can be used for visualization or subsequent system analysis and decision-making.

[0055] Simulation Implementation Example: The training and testing of the method of this invention can be performed in a typical deep learning environment. A deep learning framework based on PyTorch 1.12.0 is used, with Python 3.7 as the programming language. An NVIDIA GeForce RTX 3090 GPU is used for accelerated computation, and the training environment is configured with CUDA 11.7. During the training phase, a momentum-driven stochastic gradient descent (SGD) optimizer is used, with momentum set to 0.9 and weight decay to 0.005. The model is trained for 200 iterations, with an initial learning rate of 0.001, and a multistep learning rate decay strategy is employed. The batch size is 16, which can be set according to the GPU memory.

[0056] Corresponding to the aforementioned embodiment of an infrared target detection method based on a bidirectional receptive field attention feature network, the present invention also provides an embodiment of an infrared target detection device based on a bidirectional receptive field attention feature network.

[0057] See Figure 4 The present invention provides an infrared target detection device based on a bidirectional receptive field attention feature network, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement an infrared target detection method based on a bidirectional receptive field attention feature network as described in the above embodiment.

[0058] The present invention provides an embodiment of an infrared target detection device based on a bidirectional receptive field attention feature network. This embodiment can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 4 The diagram shown is a hardware structure diagram of any data processing-capable device, including the infrared target detection device based on a bidirectional receptive field attention feature network provided by this invention. (Except for...) Figure 4 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0059] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0060] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0061] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements an infrared target detection method based on a bidirectional receptive field attention feature network as described in the above embodiments.

[0062] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0063] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the infrared target detection method based on a bidirectional receptive field attention feature network.

[0064] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0065] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An infrared target detection method based on a bidirectional receptive field attention feature network, characterized in that, The method includes: A dataset was constructed by selecting multimodal remote sensing image data including infrared, synthetic aperture radar, and visible light, and then preprocessed. A feature encoding network is constructed by alternating receptive field attention convolutions, traditional convolutions, and C3 modules, with context aggregation performed at the network ends; data is input into the feature encoding network to obtain feature maps at different levels; The feature encoding network specifically includes: a first standard convolutional layer, a first receptive field attention convolutional layer, a first C3 convolutional layer, a second receptive field attention convolutional layer, a second C3 convolutional layer, a third receptive field attention convolutional layer, a third C3 convolutional layer, a second standard convolutional layer, and a fourth C3 convolutional layer connected in sequence. SPPF is used at the end of the network. The features output by the fourth C3 convolutional layer are processed in parallel by multiple max pooling kernels of different sizes, and the results are concatenated. Finally, the second receptive field attention convolutional layer, the third receptive field attention convolutional layer, and the SPPF-processed result are used as high-resolution, medium-resolution, and low-resolution feature maps, respectively, and input into the feature fusion network. The feature maps at different levels are passed through a feature fusion network, which includes a top-down path to pass high-level semantic information and a bottom-up path to supplement low-level spatial details. Finally, three optimized feature maps are generated and fed into three parallel detection heads, which make predictions on their respective feature maps. The top-down path specifically includes: The lowest resolution feature map output by SPPF is first passed through a receptive field attention convolutional layer for feature enhancement; The enhanced feature map is initially upsampled to match the medium-resolution feature map. The feature map after the initial upsampling is weighted and fused with the medium-resolution feature map to obtain the first fused feature; The first fusion feature is refined and enhanced using a C3 layer and a receptive field attention convolutional layer. The refined and enhanced features are then upsampled a second time to match their spatial dimensions with the high-resolution feature map. The feature map after secondary upsampling is weighted and fused with the high-resolution feature map to obtain the second fused feature; The second fusion feature is processed using the C3 layer and then input into the first detection head used to detect small targets. The bottom-up path is specifically as follows: The secondary fusion features processed by the C3 layer are then processed by the receptive field attention convolutional layer and used as the starting point of the path. The initial fusion features of the path starting point features, after downsampling and refinement enhancement, are weighted and fused to obtain the third fusion feature with medium resolution; The third fusion feature is processed by the C3 layer and then input into the second detection head used to detect medium-sized targets. The third fusion feature processed by the C3 layer is processed using a receptive field attention convolutional layer; The processed features are downsampled and then weighted and fused with the low-resolution feature map to obtain the fourth fused feature, and the medium-resolution information is added back to the low-resolution path. The fourth fusion feature is processed using C3 and then input into the third detection head used to detect large targets; Loss calculation and backpropagation are performed based on the dataset to obtain the trained feature encoding network and feature fusion network. The trained model weights are loaded, and the obtained prediction results are decoded into image coordinates and processed into detection boxes.

2. The infrared target detection method based on a bidirectional receptive field attention feature network according to claim 1, characterized in that, The preprocessing includes: The image is preprocessed using standardization. Meanwhile, to address the noise and background texture interference present in the infrared image, wavelet transform is used to decompose the image at multiple scales, separating the low-frequency contour information and high-frequency details and noise, preserving the edge features of the target, and suppressing irrelevant high-frequency noise interference. Adaptive histogram equalization is used to improve the contrast between the target and the background in the infrared image. Data augmentation was performed on the dataset: geometric transformations were used to simulate the changes in target pose, position, and size caused by different viewpoints and distances in actual remote sensing images; the brightness, contrast, and saturation of the images were randomly adjusted; and image regions were randomly cropped at different scales to generate training samples containing targets of different sizes.

3. The infrared target detection method based on a bidirectional receptive field attention feature network according to claim 1, characterized in that, The receptive field attention convolutional layer comprises three parallel sub-modules, specifically: Weight generation: The input feature map first passes through a 3×3 average pooling layer, and then is fed into a 1×1 grouped convolution; the output of this convolutional layer is shaped and passed through a Softmax activation function to generate normalized attention weights for different spatial locations within the receptive field; Feature generation: The input feature map is simultaneously fed into a 3×3 grouped convolution, and then subjected to batch normalization and ReLU activation function to generate the original spatial features; Weighting and Integration: The attention weights are multiplied element-wise with the spatial features to achieve dynamic weighting of the feature map, enabling the network to adaptively focus on more important regions. The weighted features are then passed through a standard 3×3 convolutional layer, batch normalization, and SiLU activation function to output the final enhanced features.

4. The infrared target detection method based on a bidirectional receptive field attention feature network according to claim 1, characterized in that, The three optimized feature maps are fed into three parallel detection heads, which then perform predictions on their respective feature maps. The three detection heads process the fused feature maps at different scales to adapt to targets of different sizes; each detection head is a lightweight Conv2d convolutional layer to generate the final prediction result, which includes bounding box coordinates, target confidence, and class probability.

5. The infrared target detection method based on a bidirectional receptive field attention feature network according to claim 1, characterized in that, The loss calculation and backpropagation are specifically as follows: An improved EIoU loss function is used to optimize the model's localization accuracy. Based on the CIoU loss function, the difference between width and height is calculated separately to replace the aspect ratio, directly minimizing the difference in width and height between the predicted target bounding box and the ground truth bounding box. Simultaneously, a loss function is introduced to address the imbalance between easy and difficult samples, thus achieving better localization results. The calculation formula is as follows: Where A and B represent the ground truth bounding box and the predicted bounding box, The intersection of A and B divided by the union of A and B. , Let A be the center point of B. , Let A be the width and height. , Let B be the width and height. , These are the width and height of the smallest bounding rectangle of A and B; Confidence loss Classification loss Both methods use binary cross-entropy loss, and the general formula for calculating binary cross-entropy loss is as follows: Where y is the true label, and for confidence loss Positive samples are represented by 1, and negative samples by 0; for classification loss 1 represents the target belonging to this class, and 0 represents it not belonging to this class. p is the probability predicted by the model. The total loss is the weighted sum of the EIoU location loss, confidence loss, and classification loss.

6. An infrared target detection device based on a bidirectional receptive field attention feature network, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements an infrared target detection method based on a bidirectional receptive field attention feature network as described in any one of claims 1-5.

7. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements an infrared target detection method based on a bidirectional receptive field attention feature network as described in any one of claims 1-5.

Images