Infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion

Abstract

The invention discloses an infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion, and mainly solves the problems of easy feature loss and low detection precision caused by weak target and complex background in the existing infrared small target detection technology. The implementation scheme is as follows: 1) acquiring a data set and a segmentation label; 2) constructing an infrared small target detection model; 3) constructing a loss function; 4) training an infrared small target detection model; and 5) obtaining an infrared small target detection result. According to the infrared small target detection model constructed by the invention, through a diffusion network encoder pre-trained based on a visual thermal diffusion principle, feature representation of a weak target is enhanced from the source; intelligent fusion and enhancement of multi-level features of the encoder are realized through a selective dimension fusion module deployed on jump connection; and through a selective cavity channel refining module and a selective receptive field fusion module which are embedded in the network, the high efficiency of the whole architecture is ensured while the fine extraction capability of the network to small targets is improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to an infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion. Background Technology

[0002] Infrared small target detection technology aims to identify targets of interest from infrared images using algorithms, and has significant application value in fields such as military reconnaissance and surveillance, and industrial fault detection. However, in practical applications, accurate and robust detection of infrared small targets still faces a series of technical challenges. For example, due to the long imaging distance and limitations of sensor physical conditions, infrared targets typically occupy only a very small number of pixels in the image, exhibiting tiny size, dim signal, and a severe lack of clear shape, texture, and other distinguishable geometric features. Target signals are easily submerged in complex background noise, such as cloud clutter and ground thermal radiation interference. These interference sources may exhibit brightness and size characteristics similar to real targets in local areas, resulting in extremely low image signal-to-noise ratios, which poses a significant challenge to the effective separation of targets from the background. The root cause of these problems is that existing technologies mostly rely on manually designed prior models, resulting in poor generalization ability and difficulty in achieving ideal results in variable scenarios. Furthermore, in deep convolutional neural networks, the network must perform multiple downsampling operations to obtain high-level semantic information. During this process, the shallow spatial features of the already weak small targets are easily filtered or diluted, leading to information loss in the deep network. Therefore, how to effectively suppress the loss and dilution of small target feature information during the forward propagation process of deep networks, and enhance its recognizability in complex backgrounds, is a technical problem that urgently needs to be solved in the field of infrared small target detection. Summary of the Invention

[0003] This invention aims to address the technical challenges of low target detection accuracy and poor robustness caused by the small size and dim signal of infrared small targets, as well as the "feature dilution" problem resulting from downsampling in deep networks. The invention provides an infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion. It achieves continuous high-resolution feature enhancement of weak targets in the original infrared image through a diffusion model built on the principle of visual thermal diffusion. Through a selective dimension fusion module and a selective hole channel refinement module, it achieves adaptive fusion and refinement of features of different dimensions, improving the network's feature extraction capability for small targets in complex backgrounds. Through a selective receptive field fusion module, it dynamically adjusts the weights of multi-scale features using an attention mechanism and aggregates the final feature map using feature recalibration, thereby enhancing the saliency of key target regions and suppressing background noise interference.

[0004] I. Technical Principles

[0005] Currently, most deep learning-based infrared small target detection algorithms adopt an "encoder-decoder" structure. These algorithms use an encoder to downsample the input infrared image layer by layer to extract deep semantic features, followed by a decoder to upsample and fuse these features. Skip connections are typically used to pass shallow spatial detail information from the encoder to the decoder. Finally, the detection head performs pixel-level prediction on the feature map output by the decoder to generate a target segmentation mask. In infrared small target detection applications, two major challenges remain: effectively combating the attenuation and loss of weak target features in deep networks, and avoiding excessive model complexity in pursuit of high performance. Existing general-purpose segmentation models often have limited performance when directly applied to infrared small target detection because their design does not fully consider the "small" and "dark" characteristics of infrared targets, leading to the easy loss of key information during downsampling and insufficient generalization ability. To achieve robust detection of weak targets in complex backgrounds, this invention proposes a dual-branch collaborative processing architecture. The diffusion network branch, drawing inspiration from the principle of thermal diffusion in physics, pre-enhances the input features before detection, fundamentally improving the signal-to-noise ratio of weak targets and laying the foundation for their effective transmission in subsequent networks. Meanwhile, the detection network branch, through its internal selective dimension fusion module and selective void channel refinement module, performs multi-dimensional and multi-scale refined analysis and fusion of the enhanced features, ensuring efficient utilization of target information.

[0006] II. Based on the above principles, the present invention is implemented through the following scheme:

[0007] An infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion includes the following steps:

[0008] (1) Obtaining the dataset and labels:

[0009] Obtain the infrared small target dataset and its corresponding labels;

[0010] (2) Constructing an infrared small target detection model: This model consists of a diffusion network branch and a detection network branch, wherein the encoder of the detection network branch shares weights with the encoder of the diffusion network branch; the specific construction process includes the following steps:

[0011] (2-a) Constructing the Diffusion Network Branch (DFN): This branch is used to enhance the features of the input image and generate continuous high-resolution feature representations. The DFN adopts an encoder-decoder architecture, in which a thermally diffused implicit neural representation module is embedded. The thermally diffused implicit neural representation module contains a thermally diffused layer, which uses Discrete Cosine Transform (DCT) to transform the input features. Transform to the frequency domain, then multiply the transformed features by a learnable thermal diffusivity coefficient. Calculated attenuation factor This simulates the physical process of heat diffusion, and finally transforms the features back to the spatial domain using inverse discrete cosine transform (IDCT). By simulating the heat conduction process in the physical world, noise is suppressed, thereby enhancing the target features.

[0012]

[0013] in, The input features of the heat diffusion layer are represented by the input feature map with C channels located at spatial coordinates (x,y) when time is 0. The output feature is represented by a feature map with C channels located at spatial coordinates (x, y) at time t. The thermal diffusivity is a learnable coefficient. For frequency domain coordinates, For the diffusion time step; DFN is trained independently using a self-supervised learning strategy. After training, the weights of its encoder part are frozen for use in step (2-b).

[0014] (2-b) Constructing the Detection Network Branch (DTN): This network consists of an encoder, a decoder, and multiple feature enhancement and refinement modules;

[0015] The encoder part directly uses the DFN encoder trained and frozen in step (2-a). After inputting the original infrared image, it undergoes multi-layer downsampling to obtain feature maps at four different scales. , , and Their spatial resolution decreases sequentially;

[0016] The feature enhancement and refinement function consists of a Selective Dimensional Fusion (SDPA) module, a Selective Hollow Channel Refinement (SCDR) module, and a Selective Receptive Field Fusion (SRFF) module; the three modules mentioned in this step are constructed as follows:

[0017] The Selective Dimension Fusion Module (SDPA) has three inputs: features from the same-level encoder. High-dimensional features from shallower levels of the encoder and features from deeper levels of the encoder This module first utilizes This module first generates adaptive weights α. , and The channel is evenly divided into four feature partitions. }、{ }and{ Finally, the weighted summation formula is used. The merged partition features are calculated. The module then concatenates all partition features along the channel dimension to obtain its output.

[0018] The Selective Hole Channel Refinement (SCDR) module has one input feature. This module first inputs features. Divide the channel dimension evenly into four feature heads { }; Each feature header Each part is processed through a depthwise separable dilated convolution with a different dilation rate to obtain the processing result. }; Subsequently, all processing results { The channels are interleaved and recombined, and intra-group and inter-group information is fused through pointwise convolution to obtain the output of the module. ;

[0019] The Selective Receptive Field Fusion Module (SRFF) has input features from three parallel convolutional streams. , and The three inputs have different scales; this module first sums the three inputs element-wise and generates a global feature descriptor through global average pooling; this descriptor is then processed by three parallel fully connected layers and a SoftMax activation function to generate three attention weights. , and Finally, the attention weights are weighted and summed one-to-one with the original input features. This will give you the output of the module.

[0020] The decoder consists of multiple upsampling layers; it converts the deepest feature map output by the encoder into a single layer. The input is fed into the SCDR module to obtain the refined features. ;Will Decode features are obtained after the first-stage upsampling. On the third-layer hop connection, an SDPA module is deployed to obtain enhanced hop connection features. ;Will and Feature splicing is performed to obtain the fusion result. ;Will Decode features are obtained after the second-stage upsampling. On the second-layer hop connection, an SDPA module is deployed to obtain the enhanced hop connection features. ;Will and Feature splicing is performed to obtain the fusion result. ;Will Decode features are obtained after the third-level upsampling. On the first-layer hop connection, an SDPA module is deployed to obtain the enhanced hop connection features. ;Will and Feature splicing is performed to obtain the fusion result. The fusion results , and After concatenation along the channel dimension, the data is input into the SRFF module to obtain the final output features of the decoder.

[0021] (3) Construct the loss function:

[0022] This invention relates to two independent training phases, corresponding to the diffusion network (DFN) and the detection network (DTN) respectively, thus constructing two independent loss functions:

[0023] (3-a) Constructing the loss function of the diffusion network (DFN):

[0024] This loss function is used in the first stage of training the diffusion network (DFN). Its goal is to optimize the network to ensure it can accurately predict added noise from images with added Gaussian noise, thereby learning the ability to recover high-resolution images. The loss function is constructed as follows:

[0025]

[0026] in, This indicates the loss of the diffusion network; This represents the mathematical expectation operation, which averages the loss values ​​of all sampled data within a training batch during model training; the diffusion model aims to obtain high-resolution images through a series of denoising operations. ; For time step Noisy images, This represents the variance of the Gaussian noise added in each iteration; This represents the real Gaussian noise added to the high-resolution image during training; This indicates that the diffusion network is effective in a given low-resolution image. Time step Noisy images and variance with added Gaussian noise The noise predicted under the given conditions; Represents true Gaussian noise Noise predicted by the network The loss function optimizes the network by minimizing the average absolute error between the real noise and the predicted noise.

[0027] (3-b) Constructing the loss function for the Detection Network (DTN):

[0028] This loss function is used for the second stage of end-to-end training of the Detection Network (DTN). It is a composite loss function designed to ensure accurate depiction of infrared target boundaries while preserving spatial details of small targets in class imbalance cases by introducing Dice loss. The composite loss function is constructed as follows:

[0029]

[0030] in, This represents the total loss of the detection network; This represents the binary cross-entropy loss function, used for accurate classification of each pixel; λ represents the Dice loss function, used to address the problem of a severe imbalance in the number of pixels between the target and the background; λ is a hyperparameter that balances the two losses, with an optimal value of 0.8.

[0031] (4) Training the infrared small target detection model:

[0032] The infrared small target detection model constructed in step (2) is trained using the dataset obtained in step (1); the error between the predicted result output by the model and the label is calculated using the loss function constructed in step (3); the Adam algorithm is used as the optimizer to update the model parameters during the training process until the loss no longer decreases, and the trained infrared small target detection model is obtained.

[0033] (5) Infrared small target detection:

[0034] The infrared image to be detected is input into the trained infrared small target detection model, and the binarized image output by the model is the final infrared small target detection result.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] (1) By sharing a pre-trained and frozen diffusion network encoder, the present invention provides a high-quality continuous high-resolution feature representation for the detection network, enhances weak targets, and effectively solves the problem that infrared small target features are easily lost in deep networks.

[0037] (2) The present invention constructs a selective dimension fusion module (SDPA) deployed on skip connections, which can intelligently fuse features from multiple levels of the encoder and greatly enhance the information of the features before passing them to the decoder, thus significantly improving the network's ability to make comprehensive use of spatial details and semantic context.

[0038] (3) In this invention, a Selective Hollow Channel Refinement (SCDR) module and a Selective Receptive Field Fusion (SRFF) module are constructed in the network respectively. Through multi-scale feature extraction and adaptive weighting, the network’s ability to distinguish small targets from complex backgrounds is further improved, and the salience of the final output features is enhanced. Attached Figure Description

[0039] Figure 1 Structure diagram of the infrared small target detection model according to an embodiment of the present invention;

[0040] Figure 2 A structural diagram of the heat-diffusion implicit neural representation module according to an embodiment of the present invention;

[0041] Figure 3 Structural diagram of the Selective Dimension Fusion Module SDPA according to an embodiment of the present invention;

[0042] Figure 4 Structural diagram of the Selective Void Channel Refinement Module (SCDR) according to an embodiment of the present invention;

[0043] Figure 5 Structural diagram of the Selective Receptive Field Fusion Module (SRFF) according to an embodiment of the present invention;

[0044] Figure 6 A comparison chart of the detection results of the infrared small target detection method according to the present invention with the detection results of other methods. Detailed Implementation

[0045] The specific embodiments of the present invention are described below:

[0046] Example 1

[0047] The specific steps of an infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion are as follows:

[0048] Step 1, Obtain the dataset and labels:

[0049] Obtain the infrared small target dataset and its corresponding labels;

[0050] Step 2, constructing an infrared small target detection model: This model consists of a diffusion network branch and a detection network branch, where the encoder of the detection network branch shares weights with the encoder of the diffusion network branch; the specific construction process includes the following steps:

[0051] Figure 1 The diagram shown illustrates the overall framework of an embodiment of the present invention, demonstrating the collaborative relationship between the diffusion network branch (DFN) and the detection network branch (DTN). The specific steps are as follows:

[0052] (2-a) Constructing the Diffusion Network (DFN) branch: This branch is used to enhance the features of the input image and generate continuous high-resolution feature representations. The DFN adopts an encoder-decoder architecture, in which a thermal diffusion implicit neural representation module is embedded. This module is constructed based on the thermal diffusion equation and discrete cosine transform and is used to learn continuous high-resolution features of the input image. The DFN is trained independently through a self-supervised learning strategy. After training, the weights of its encoder part are frozen for use in step (2-b).

[0053] Figure 2 The diagram shown is a structural diagram of the heat-diffusion implicit neural representation module according to an embodiment of the present invention. The specific steps are as follows:

[0054] The heat-diffusion implicit neural representation module takes discrete feature maps as input to generate continuous high-resolution feature representations. The core of this module is a heat-diffusion layer, which first obtains continuous feature representations through implicit neural representations and layer normalization, and then constructs preliminary feature representations using a 3×3 depthwise separable convolution (DWConv). Subsequently, the features are transformed to the frequency domain using discrete cosine transform, and then multiplied by a learnable heat-diffusion coefficient. Calculated attenuation factor To simulate the physical thermal diffusion process, the features are transformed back to the spatial domain through inverse discrete cosine transform, resulting in a feature map enhanced by physical principles. Finally, continuous feature representations are output through layer normalization and feedforward neural network FFN.

[0055] (2-b) Constructing the Detection Network Branch (DTN): This network consists of an encoder, a decoder, and multiple feature enhancement and refinement modules;

[0056] The encoder part directly uses the DFN encoder trained and frozen in step (2-a). After inputting the original infrared image, it undergoes multi-layer downsampling to obtain feature maps at four different scales. , , and Their spatial resolution decreases sequentially;

[0057] Figure 3 The diagram shown is a structural diagram of the Selective Dimension Fusion Module (SDPA) according to an embodiment of the present invention, and its specific construction is as follows:

[0058] The Selective Dimension Fusion (SDPA) module is deployed on the skip connection path from encoder to decoder. It has three inputs, features from the same-level encoder. High-dimensional features from shallower levels of the encoder and features from deeper levels of the encoder This module first uses convolution and bilinear interpolation operations to... and Size and Alignment was achieved, with convolution kernels of 3×3 and 1×1 respectively; subsequently, the three features were uniformly divided into four feature partitions along the channel dimension. }、{ }and{ }; Utilize the current layer partition features { The adaptive weights α are generated using the Sigmoid activation function; then, they are calculated using the weighted summation formula. The merged partition features are calculated. Finally, all partition features { The features are concatenated along the channel dimension and processed through a convolutional block consisting of pointwise convolution PWConv, batch normalization, and ReLU activation function to obtain the output features of this module.

[0059] Figure 4 The diagram shown is a structural diagram of the Selective Void Channel Refinement Module (SCDR) according to an embodiment of the present invention, and its specific construction is as follows:

[0060] The Selective Hollow Channel Refinement (SCDR) module is located at the end of the encoder path and has one input feature. This module first divides the input features evenly along the channel dimension into four feature heads. Each feature head is processed by a depthwise separable dilated convolution DDWConv with different dilation rates (preferably 1, 6, 12, or 18). Then, all channels processed by batch normalization and ReLU activation are interleaved and recombined, and intra- and inter-group information is fused using pointwise convolution PWConv to finally obtain the output feature of this module. ;

[0061] Figure 5 The diagram shown is a structural diagram of the Selective Receptive Field Fusion (SRFF) module according to an embodiment of the present invention, and its specific construction is as follows:

[0062] The Selective Receptive Field Fusion (SRFF) module is located at the end of the decoder path and has input features from three parallel convolutional streams. , and This module first sums the three inputs element-wise and performs global average pooling, then reduces the number of channels to one-eighth of the original number using a convolutional Conv operation. Subsequently, it generates three attention weights through parallel convolutional Conv operations and a SoftMax activation function. , and Finally, the attention weights are weighted and summed one-to-one with the original input features. This will give you the output of the module.

[0063] Will Decode features are obtained after the first-stage upsampling. On the third-layer hop connection, an SDPA module is deployed to obtain enhanced hop connection features. ;Will and Feature splicing is performed to obtain the fusion result. ;Will Decode features are obtained after the second-stage upsampling. On the second-layer hop connection, an SDPA module is deployed to obtain the enhanced hop connection features. ;Will and Feature splicing is performed to obtain the fusion result. ;Will Decode features are obtained after the third-level upsampling. On the first-layer hop connection, an SDPA module is deployed to obtain the enhanced hop connection features. ;Will and Feature splicing is performed to obtain the fusion result. The fusion results , and After concatenation along the channel dimension, the data is input into the SRFF module to obtain the final output features of the decoder.

[0064] Step 3, construct the loss function:

[0065] This invention relates to two independent training phases, corresponding to the diffusion network (DFN) and the detection network (DTN) respectively, thus constructing two independent loss functions:

[0066] (3-a) Constructing the loss function of the diffusion network (DFN):

[0067] This loss function is used in the first stage of training the diffusion network (DFN). Its goal is to optimize the network to ensure it can accurately predict added noise from images with added Gaussian noise, thereby learning the ability to recover high-resolution images. The loss function is constructed as follows:

[0068]

[0069] in, This indicates the loss of the diffusion network; This represents the mathematical expectation operation, which averages the loss values ​​of all sampled data within a training batch during model training; the diffusion model aims to obtain high-resolution images through a series of denoising operations. ; For time steps Noisy images, This represents the variance of the Gaussian noise added in each iteration; This represents the real Gaussian noise added to the high-resolution image during training; This indicates that the diffusion network is effective in a given low-resolution image. Time step Noisy images and variance with added Gaussian noise The noise predicted under the given conditions; Represents true Gaussian noise Noise predicted by the network The loss function optimizes the network by minimizing the average absolute error between the real noise and the predicted noise.

[0070] (3-b) Constructing the loss function for the Detection Network (DTN):

[0071] This loss function is used for the second stage of end-to-end training of the Detection Network (DTN). It is a composite loss function designed to ensure accurate depiction of infrared target boundaries while preserving spatial details of small targets in class imbalance cases by introducing Dice loss. The composite loss function is constructed as follows:

[0072]

[0073] in, This represents the total loss of the detection network; This represents the binary cross-entropy loss function, used for accurate classification of each pixel; λ represents the Dice loss function, used to address the problem of a severe imbalance in the number of pixels between the target and the background; λ is a hyperparameter that balances the two losses, with an optimal value of 0.8.

[0074] Step 4, train the infrared small target detection model:

[0075] The infrared small target detection model constructed in step (2) is trained using the dataset obtained in step (1); the error between the predicted result output by the model and the label is calculated using the loss function constructed in step (3); the Adam algorithm is used as the optimizer to update the model parameters during the training process until the loss no longer decreases, and the trained infrared small target detection model is obtained.

[0076] Step 5, Infrared small target detection:

[0077] The infrared image to be detected is input into the trained infrared small target detection model, and the binarized image output by the model is the final infrared small target detection result.

[0078] Example 2

[0079] Infrared small target detection experiments were conducted on the NUST-SIRST infrared small target dataset using the method described in Example 1. This experiment was based on the PyTorch 1.11.0 framework and was trained and tested on a server equipped with an Intel(R) Xeon(R) Platinum 8362 CPU (@ 2.80GHz) and an NVIDIA GeForce RTX 3090 (24GB) graphics card.

[0080] This example uses four metrics—mIoU, F1 score, inference time, and number of parameters—to conduct an experimental comparison of six testing methods (RDIAN, MLCL, LSPM, DNA, MTU, and UIU) with the method of this invention on the NUST-SIRST dataset.

[0081] The comparison results are shown in Table 1. It can be seen that, compared with other methods, the present invention can obtain accurate detection results on the NUST-SIRST dataset and achieves the best performance in terms of evaluation metrics mIoU, F1_score and inference time.

[0082] Figure 6 The image shows a comparison of the infrared small target detection results of the embodiment of the present invention with those of other methods. The results demonstrate that the model designed in this invention can effectively handle various challenging scenarios, including those with strong background noise interference. Figure 6 Column 1), a scene where the target is dim and small ( Figure 6 Column 4) and scenarios containing multiple targets ( Figure 6 (Column 6). Compared with other methods, the detection results obtained by this invention are not only accurate in positioning, but also the segmented target contours are more complete and closer to the real labels, while effectively suppressing the occurrence of false detections.

[0083] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the structure and principles of the present invention should be covered within the protection scope of the present invention.

[0084] Table 1

[0085] Methods mIoU F1_score Inference time (s) Parameter number (M) RDIAN 0.4076 0.5054 15.87 0.22 MLCL 0.4296 0.5782 46.89 0.55 LSPM 0.4225 0.5419 315.78 31.58 DNA 0.4387 0.5914 60.81 4.70 MTU 0.4275 0.5873 18.68 12.75 UIU 0.4423 0.6076 39.61 50.54 This invention 0.4492 0.6295 12.33 146.82

Claims

1. A method for detecting small infrared targets based on the fusion of visual thermal diffusion and selective multi-scale features, characterized in that... Includes the following steps: (1) Obtaining the dataset and labels: Obtain the infrared small target dataset and its corresponding labels; (2) Constructing an infrared small target detection model: This model consists of a diffusion network branch and a detection network branch, wherein the encoder of the detection network branch shares weights with the encoder of the diffusion network branch; the specific construction process includes the following steps: (2-a) Constructing the Diffusion Network Branch (DFN): This branch is used to enhance the features of the input image and generate continuous high-resolution feature representations. The DFN adopts an encoder-decoder architecture, in which a thermally diffused implicit neural representation module is embedded. The thermally diffused implicit neural representation module contains a thermally diffused layer, which uses Discrete Cosine Transform (DCT) to transform the input features. Transform to the frequency domain, then multiply the transformed features by a learnable thermal diffusivity coefficient. Calculated attenuation factor This simulates the physical process of heat diffusion, and finally transforms the features back to the spatial domain using inverse discrete cosine transform (IDCT); by simulating the heat conduction process in the physical world, noise is suppressed, thereby enhancing the target features. in, The input features of the heat diffusion layer are represented by the input feature map with C channels located at spatial coordinates (x,y) when time is 0. The output feature is represented by a feature map with C channels located at spatial coordinates (x, y) at time t. The thermal diffusivity is a learnable coefficient. For frequency domain coordinates, For the diffusion time step; DFN is trained independently using a self-supervised learning strategy. After training, the weights of its encoder part are frozen for use in step (2-b). (2-b) Constructing the Detection Network Branch (DTN): This network consists of an encoder, a decoder, and multiple feature enhancement and refinement modules; The encoder part directly uses the DFN encoder trained and frozen in step (2-a). After inputting the original infrared image, it undergoes multi-layer downsampling to obtain feature maps at four different scales. , , and Their spatial resolution decreases sequentially; The feature enhancement and refinement function consists of a Selective Dimensional Fusion (SDPA) module, a Selective Hollow Channel Refinement (SCDR) module, and a Selective Receptive Field Fusion (SRFF) module; the three modules mentioned in this step are constructed as follows: The Selective Dimension Fusion Module (SDPA) has three inputs: features from the same-level encoder. High-dimensional features from shallower levels of the encoder and features from deeper levels of the encoder This module first utilizes This module first generates adaptive weights α. , and The channel is evenly divided into four feature partitions. }、{ }and{ Finally, the weighted summation formula is used. The merged partition features are calculated. The module then concatenates all partition features along the channel dimension to obtain its output. The Selective Hole Channel Refinement (SCDR) module has one input feature. This module first inputs features. Divide the channel dimension evenly into four feature heads { }; Each feature header Each part is processed through a depthwise separable dilated convolution with a different dilation rate to obtain the processing result. }; Subsequently, all processing results { The channels are interleaved and recombined, and intra-group and inter-group information is fused through pointwise convolution to obtain the output of the module. ; The Selective Receptive Field Fusion Module (SRFF) has input features from three parallel convolutional streams. , and The three inputs have different scales; this module first sums the three inputs element-wise and generates a global feature descriptor through global average pooling; this descriptor is then processed by three parallel fully connected layers and a SoftMax activation function to generate three attention weights. , and Finally, the attention weights are weighted and summed one-to-one with the original input features. This will give you the output of the module. The decoder consists of multiple upsampling layers; it converts the deepest feature map output by the encoder into a single layer. The input is fed into the SCDR module to obtain the refined features. ;Will Decode features are obtained after the first-stage upsampling. On the third-layer hop connection, an SDPA module is deployed to obtain enhanced hop connection features. ;Will and Feature concatenation is performed to obtain the fusion result. ;Will Decode features are obtained after the second-stage upsampling. On the second-layer hop connection, an SDPA module is deployed to obtain the enhanced hop connection features. ;Will and Feature concatenation is performed to obtain the fusion result. ;Will Decode features are obtained after the third-level upsampling. On the first-layer hop connection, an SDPA module is deployed to obtain the enhanced hop connection features. ;Will and Feature concatenation is performed to obtain the fusion result. The fusion results , and The summation is then fed into the SRFF module to obtain the final output features of the decoder; (3) Construct the loss function: This invention relates to two independent training phases, corresponding to the diffusion network (DFN) and the detection network (DTN) respectively, thus constructing two independent loss functions: (3-a) Constructing the loss function of the diffusion network (DFN): This loss function is used in the first stage of training the diffusion network (DFN). Its goal is to optimize the network to ensure it can accurately predict added noise from images with added Gaussian noise, thereby learning the ability to recover high-resolution images. The loss function is constructed as follows: in, This indicates the loss of the diffusion network; This represents the mathematical expectation operation, which averages the loss values ​​of all sampled data within a training batch during model training; the diffusion model aims to obtain high-resolution images through a series of denoising operations. ; For time step Noisy images, This represents the variance of the Gaussian noise added in each iteration; This represents the real Gaussian noise added to the high-resolution image during training; This indicates that the diffusion network is effective in a given low-resolution image. Time step Noisy images and variance with added Gaussian noise The noise predicted under the given conditions; Represents true Gaussian noise Noise predicted by the network The loss function optimizes the network by minimizing the mean absolute error between the real noise and the predicted noise. (3-b) Constructing the loss function for the Detection Network (DTN): This loss function is used for the second stage of end-to-end training of the Detection Network (DTN). It is a composite loss function designed to ensure accurate depiction of infrared target boundaries while preserving spatial details of small targets in class imbalance cases by introducing Dice loss. The composite loss function is constructed as follows: in, This represents the total loss of the detection network; This represents the binary cross-entropy loss function, used for accurate classification of each pixel; λ represents the Dice loss function, used to address the problem of a severe imbalance in the number of pixels between the target and the background; λ is a hyperparameter that balances the two losses, with an optimal value of 0.

8. (4) Training the infrared small target detection model: The infrared small target detection model constructed in step (2) is trained using the dataset obtained in step (1); the error between the predicted result output by the model and the label is calculated using the loss function constructed in step (3); the Adam algorithm is used as the optimizer to update the model parameters during the training process until the loss no longer decreases, and the trained infrared small target detection model is obtained. (5) Infrared small target detection: The infrared image to be detected is input into the trained infrared small target detection model, and the binarized image output by the model is the final infrared small target detection result.

2. The infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion as described in claim 1, characterized in that, The diffusion network branch (DFN) in step (2-a) enhances the target features through a thermal diffusion implicit neural expression module; the thermal diffusion implicit neural expression module contains a thermal diffusion layer, which uses discrete cosine transform (DCT) to transform the input features. Transform to the frequency domain, then multiply the transformed features by a learnable thermal diffusivity coefficient. Calculated attenuation factor To simulate the physical thermal diffusion process, the features are transformed back to the spatial domain through inverse discrete cosine transform (IDCT) to suppress noise. Before DCT processing, a feature extraction module consisting of layer normalization and 3×3 depthwise separable convolutions (DWConv) is used to extract local features and stabilize the data distribution in a lightweight manner, providing high-quality input for subsequent simulated thermal diffusion processing.

3. The infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion as described in claim 1, characterized in that, The Selective Hollow Channel Refinement (SCDR) module in step (2-b) divides the input features into multiple parallel feature heads along the channel dimension. Each feature head is processed by a depthwise separable dilated convolution with different dilation rates (preferably 1, 6, 12, and 18). By capturing spatial features of different receptive field sizes in parallel, the difference between small infrared targets and complex backgrounds can be modeled more precisely.

4. The infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion as described in claim 1, characterized in that, The Selective Hollow Channel Refinement (SCDR) module in step (2-b) performs dilated convolution on multiple feature heads, interleaves the channels of all feature heads to form new feature groups, and then performs intra-group and inter-group information fusion on the new feature groups through pointwise convolution, thereby enhancing the information interaction between features at different scales to improve feature representation capabilities and achieving efficient feature fusion in a lightweight manner.

5. The infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion as described in claim 1, characterized in that, The Selective Dimensional Fusion (SDPA) module in step (2-b) aligns the three input features from the current level, shallower levels, and deeper levels of the encoder in terms of size and partitions them in the channel dimension. It uses the partitioned features of the current level to generate a set of adaptive attention weights α through the Sigmoid activation function. The attention weights α are then used to fuse features by weighted summation. This effectively solves the technical problem of high-dimensional features losing target information and low-dimensional features lacking context due to multiple downsampling by dynamically balancing high-resolution detail information and high-level semantic context information.

6. The infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion as described in claim 1, characterized in that, The Selective Receptive Field Fusion (SRFF) module in step (2-b) generates a global feature descriptor by summing multiple parallel input features element-wise and performing global average pooling. It then uses a channel-reduction convolutional layer to reduce the number of feature channels in the global feature descriptor to one-eighth of the original number of channels, improving computational efficiency. Subsequently, it generates multiple adaptive attention weights through parallel upsampling layers and the SoftMax activation function to dynamically adjust the weights of features at different scales, thereby calibrating and fusing features. This achieves dynamic integration of multi-scale contextual information and enhances the feature representation of small targets while maintaining a lightweight model.

7. The infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion as described in claim 1, characterized in that, In step (3), the composite loss function has a preferred value of 0.8 for the hyperparameter λ.

8. The infrared small target detection method based on visual thermal diffusion and selective multi-scale feature fusion as described in claim 1, characterized in that, The encoder in step (2-a) has its weights frozen during the training process in step (4) and is not updated.

Images