An IPI guiding feature decoupling-based cross-domain infrared small target detection method

Abstract

The application discloses a cross-domain infrared small target detection method based on IPI guided feature decoupling. The method comprises the following steps: acquiring a source domain infrared image and labeling a binary real mask to serve as a training set; constructing an IPI guided feature decoupling module comprising a compression up-sampling, a feature block construction, a robust hyperparameter estimation, an iterative decoupling and a target reconstruction supervision unit, and introducing the decoder of an infrared small target detection network, inputting the training set into the network, removing the IPI guided feature decoupling module after training, inputting a target domain infrared image into the trained network, and outputting a cross-domain small target detection result after processing. The application aims at the problem that the existing detection network has insufficient generalization ability in cross-sensor, cross-scene and cross-dataset tests, introduces the low-rank background and sparse target prior in the IPI into the network decoder to constrain the network to learn a stable background-target structure representation, and can improve the robustness and generalization performance of the network in unknown domains without increasing the inference overhead.

Description

Technical Field

[0001] This invention relates to an infrared small target detection method, which relates to the fields of infrared image processing, target detection and cross-domain generalization technology, and specifically to a cross-domain infrared small target detection method based on IPI-guided feature decoupling. Background Technology

[0002] Infrared small target detection has significant application value in scenarios such as maritime search and rescue, early warning and surveillance, remote sensing, and target discovery on unmanned platforms. Due to the extremely small target scale, weak texture information, and low signal-to-noise ratio, and the significant differences between imaging platforms, sensor types, scene backgrounds, and environmental conditions in different data domains, the detection performance of existing detection models often declines significantly in unknown domains.

[0003] Simulated data generation and data augmentation can improve the diversity of training, but they either fail to realistically simulate the features of real infrared imaging or are limited to perturbations of the source distribution. Domain generalization is better suited for this scenario because it aims to learn domain-invariant representations without requiring test data to follow the source distribution. However, general domain generalization methods are not specifically designed for infrared small target detection, as tiny and poorly textured targets in infrared small target detection are easily confused with background noise.

[0004] Existing deep learning-based infrared small target detection methods typically employ pixel-level segmentation loss as an optimization strategy, achieving high detection accuracy under intra-domain testing conditions. However, the learned features are easily dependent on the statistical properties of the source domain background, leading to missed detections and false detections in cross-domain scenarios. Conversely, traditional methods provide valuable physical priors. However, methods based on filtering and local contrast rely primarily on local saliency, assuming that the infrared small target produces a stronger response than its surrounding environment. While these local priors are simple and transferable, they are highly sensitive to background clutter, edges, and noise that may lead to similar responses. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides a cross-domain infrared small target detection method based on IPI-guided feature decoupling. The purpose of this invention is to solve the technical problems of existing infrared small target detection networks lacking stable structural constraints in cross-domain scenes, having insufficient generalization ability, and the difficulty in effectively integrating traditional IPI priors into the deep feature learning process. This invention does not reconstruct a new infrared small target detection network architecture, but rather introduces a pluggable training guidance module into the existing encoder-decoder detection network. By introducing the low-rank background and sparse target priors from the infrared block image model IPI into the decoding feature space of the encoder-decoder detection network, the network is constrained to learn a stable background-target structure representation. This module is deployed throughout the entire decoder stage. to By imposing low-rank background and sparse target structural constraints on multi-scale decoding features to internalize IPI priors and completely removing them during the inference stage, the robustness and generalization performance of the infrared small target detection model in the unknown domain can be improved without increasing additional inference overhead.

[0006] The technical solution adopted in this invention is: The cross-domain infrared small target detection method based on IPI-guided feature decoupling of the present invention includes: Step S1: Acquire several source domain infrared images and label them with binary ground truth masks (GT) to use as the training set.

[0007] Step S2: Construct an Infrared Patch-Image Guided Feature Decoupling Module (IGFD) comprising a compression upsampling unit, a feature block construction unit, a robust hyperparameter estimation unit, an iterative decoupling unit, and a target reconstruction supervision unit. Introduce the IGFD module into the decoder of the infrared small target detection network and construct a total loss function for the infrared small target detection network, including the main network segmentation loss and the IGFD module auxiliary loss. Input the training set into the infrared small target detection network for training until the total loss function converges. Remove the IGFD module to obtain the trained infrared small target detection network, i.e., only the original main detection branch of the network is retained for inference on the infrared image of the target domain to be detected.

[0008] Step S3: Input the infrared image of the target domain to be detected into the trained infrared small target detection network, process it, and output the detection results of cross-domain small targets in the infrared image of the target domain.

[0009] In step S2, the infrared small target detection network includes an encoder, a decoder, and a detection head connected in sequence. The decoder includes L decoding layers. Each decoding layer receives the output of the encoder, processes it, and outputs its own decoded features. The detection head outputs the main prediction mask. During the training phase of the infrared small target detection network, an IPI-guided feature decoupling module (IGFD) is introduced after each decoding layer but is not connected to the next decoding layer. For each decoding layer, the output decoded features are first processed by a compression upsampling unit to output a single-channel feature map. Then, they are input into the feature block construction unit and the robust hyperparameter estimation unit, respectively, and processed to output a feature block matrix and an iterative decoupling hyperparameter. These are then input into the iterative decoupling unit, processed, and output a sparse target feature matrix. Finally, they are input into the target reconstruction supervision unit, processed, and output an auxiliary prediction mask. A total loss function is constructed based on the auxiliary prediction mask, the binary true mask, and the main prediction mask. The decoder internalizes the IPI prior through the auxiliary prediction mask, and the binary true mask supervises the auxiliary prediction mask and the main prediction mask.

[0010] In step S2, when the decoded feature input is processed in the compression and upsampling unit CUB, feature compression is performed first, followed by upsampling, and then a single-channel feature map is output. .

[0011] In step S2, the single-channel feature map When processing the input feature block construction unit, a size of [size missing] is used. Step size is Sliding window for single-channel feature maps Perform local patch extraction (Unfold), then vectorize each local patch and stack them column-wise to obtain the feature block matrix. , , , , ,in, This represents the low-rank background feature matrix. Let N represent the sparse target feature matrix, and let N represent the total number of extracted local patch blocks.

[0012] In step S2, the single-channel feature map When the input is processed in the Robust Hyperparameter Estimation Unit (RHE), it is sequentially processed by Global Average Pooling (GAP), Multilayer Perceptron (MLP), and the Sigmoid activation function before outputting iterative decoupling hyperparameters. These iterative decoupling hyperparameters include the target sparse regularization parameters obtained in each iteration during the training phase. and augmented Lagrange penalty parameters .

[0013] In step S2, the feature block matrix When the hyperparameters of the iterative decoupling process are input together into the iterative decoupling unit for processing, a constrained optimization problem is established. as follows:

[0014] in, and Representing the feature block matrix respectively The low-rank background feature matrix and sparse target feature matrix; Represents the Lagrange multiplier matrix; Represents the nuclear norm. express Norm, Represents the inner product of matrices. This represents the Frobenius norm.

[0015] Then, the Inexact Augmented Lagrange Multiplier (IALM) method is used to perform... The next iteration updates explicitly separate the low-rank background feature matrix. Sparse target feature matrix In the In this iteration, the sparse target feature matrix is ​​processed based on the Alternating Direction Method of Multipliers (ADMM). The following updates have been made:

[0016]

[0017]

[0018] in, Indicates the first The sparse target feature matrix of the next iteration; Indicates based on parameters The soft threshold operator, , and They represent the first The target sparse regularization parameter and the augmented Lagrange penalty parameter for the next iteration; Indicates the first The low-rank background feature matrix of the next iteration; and They represent the first Subsequent The Lagrange multiplier matrix of the next iteration; Indicates based on parameters The Singular Value Thresholding (SVT) operator is used in the background feature matrix update. .

[0019] go through After several iterations, the final output is the decoupled sparse target feature matrix. .

[0020] In step S2, the sparse target feature matrix During the input target reconstruction supervision unit, feature reconstruction (Fold) and average overlapping region (AOR) processing are performed sequentially to reconstruct a spatial domain target feature map. Then, after processing by the Target Refinement Convolution (TRC) unit, an auxiliary prediction mask is output. .

[0021] In step S2, the total loss function Specifically as follows:

[0022]

[0023] in, and These represent the main network segmentation loss and the IGFFD module auxiliary loss, respectively. This represents the main prediction mask output by the detection head; Represents a binary true mask; α Indicates the loss balance coefficient; Indicates Soft-IoU loss; The decoder's first l The auxiliary prediction mask obtained from the layer decoding layer.

[0024] The electronic device of the present invention includes: a memory and a processor coupled to each other, wherein the memory stores program data, and the processor invokes the program data to execute the method described above.

[0025] The present invention provides a computer-readable storage medium having program data stored thereon, which, when executed by a processor, implements the method described above.

[0026] The Infrared Patch-Image (IPI) model describes infrared small target detection from the perspective of background-target decomposition. By organizing local patches into a patch image matrix, IPI separates common redundant backgrounds from rare target responses, modeling them as low-rank and sparse components, respectively. Compared to priors based on local responses, this low-rank sparse formulation captures a more intrinsic structure of the infrared scene, thus providing more robust cues across different domains.

[0027] This invention integrates IPI priors into a deep learning detector to improve the model's generalization performance. Addressing the insufficient generalization ability of existing deep infrared small target detection networks in cross-sensor, cross-scene, and cross-dataset tests, this invention introduces low-rank background and sparse target priors from IPI into the decoder of the infrared small target detection network to constrain the network to learn a stable background-target structure representation. During the training phase, the module decouples the low-rank background and sparse target from the decoded features and uses target components for auxiliary supervision, thereby internalizing the IPI priors into the network parameters. During the inference phase, the module is completely removed, retaining only the original detection branch outputting the detection results. Therefore, this invention can improve the robustness and generalization performance of the infrared small target detection model in unknown domains without increasing inference overhead.

[0028] The beneficial effects of this invention are: 1. Improve cross-domain detection robustness by modeling feature structures based on IPI priors: To address the problem that existing deep infrared small target detection networks mainly rely on pixel-level segmentation loss and are prone to learning source domain-specific background textures and noise patterns, this invention introduces the low-rank background prior and sparse target prior from the IPI model into the decoder's multi-scale feature learning process to constrain the feature learning process. This structural constraint can suppress domain-specific background interference from the feature composition level, so that the network no longer relies solely on the apparent distribution of the source domain for discrimination, but learns a background-target structure representation that is more consistent with the imaging mechanism of infrared small targets. This effectively reduces missed detections and false detections in unknown domain scenes and improves the model's detection robustness under cross-sensor, cross-dataset, and complex background change conditions.

[0029] 2. Multi-stage pluggable training supervision, maintaining zero additional overhead in the inference stage: The IGFD module proposed in this invention can be pluggably deployed on the decoder. to In multiple stages, low-rank sparse structure constraints are simultaneously applied to deep semantic features and shallow high-resolution features through multi-scale assisted supervision, improving the consistency of decoded features at different scales. Unlike methods that require continuous feature alignment, decomposition, or additional branch computation during the inference phase, the IGFD module of this invention is only enabled during the training phase; after training, the IGFD module can be completely removed, and only the main detection branch of the original infrared small target detection network is retained during the inference phase. Therefore, this invention improves cross-domain generalization performance without increasing the number of additional parameters, computational load, or inference latency, making it suitable for deployment on real-time infrared sensing and resource-constrained platforms.

[0030] In summary, by introducing IPI physical priors into the deep decoding feature space, the infrared small target detection network can internalize low-rank background and sparse target structure priors during training, effectively improving the target detection accuracy and robustness in unseen domain scenarios, while also possessing advantages such as flexible structure, pluggable deployment, and efficient inference. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the structure of the encoder-decoder type infrared small target detection network embedded in the IGFD module of the present invention; Figure 3 This is a schematic diagram illustrating the cross-domain detection effect of the method of the present invention when applied to different infrared small target detection networks. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1 As shown, the cross-domain infrared small target detection method based on IPI-guided feature decoupling of the present invention is as follows: First, several source domain infrared images are acquired and labeled with binary ground truth masks (GT) to form the training set.

[0034] Then, an IPI-guided feature decoupling module (IGFD) is constructed, comprising a compression upsampling unit, a feature block construction unit, a robust hyperparameter estimation unit, an iterative decoupling unit, and a target reconstruction supervision unit. The IGFD module is then introduced into the decoder of the infrared small target detection network, and a total loss function for the infrared small target detection network, including the main network segmentation loss and the IGFD module auxiliary loss, is constructed. The training set is input into the infrared small target detection network for training until the total loss function converges. The IGFD module is then removed, resulting in the trained infrared small target detection network, which retains only the original main detection branch for inference on the infrared image of the target domain to be detected.

[0035] like Figure 2 As shown, the infrared small target detection network includes an encoder, a decoder, and a detection head connected in sequence. Specifically, it can employ SCTransNet, DNANet, UIUNet, or other U-Net-type encoder-decoder infrared small target detection networks, without limiting the specific backbone structure. The decoder includes L decoding layers, each receiving and processing the encoder's output before outputting its own decoded features. The encoder includes L encoding layers, confirming the transmission relationship between the encoder and decoder based on the adopted infrared small target detection network. The detection head outputs the master prediction mask. Specifically, the input infrared image is denoted as... The corresponding binary true mask is denoted as Encoder Used for extracting multi-scale semantic features layer by layer, decoder The main detection head is used to progressively restore spatial resolution and fuse semantic and spatial information at different levels. Used to output the main target prediction map .

[0036]

[0037]

[0038]

[0039] in, Indicates the first The encoded features output by each coding layer Indicates the first Decoding features output by each decoding layer This indicates the total number of decoding layers. (This refers to the total number of stages in the decoder.) to Deploy the IPI-guided feature decoupling module IGFD.

[0040] During the training phase of the infrared small target detection network, an IPI-guided feature decoupling module (IGFD) is introduced after each decoding layer, but it is not connected to the next decoding layer. For each decoding layer, the output decoded features are first processed by the compression upsampling unit to output a single-channel feature map. Then, they are respectively input into the feature block construction unit and the robust hyperparameter estimation unit for processing, and the output features block matrix and iterative decoupling hyperparameters are output respectively. Then, they are jointly input into the iterative decoupling unit for processing, and the output is a sparse target feature matrix. Then, they are input into the target reconstruction supervision unit for processing, and the output is an auxiliary prediction mask. The total loss function is constructed based on the auxiliary prediction mask, the binary true mask, and the main prediction mask. The decoder internalizes the IPI prior through the auxiliary prediction mask, and the binary true mask supervises the auxiliary prediction mask and the main prediction mask.

[0041] The IGFD module is embedded only in the decoder of the infrared small target detection network during the training phase to impose low-rank background and sparse target structure constraints on the decoded features to internalize the IPI prior; during the inference phase, the IGFD module is removed so as not to add additional inference overhead.

[0042] When the decoded feature input is processed in the compression and upsampling unit (CUB), feature compression is performed first, followed by upsampling, and then a single-channel feature map is output. .

[0043] The compressed upsampling unit (CUB) consists of a 3×3 convolutional layer (conv), an activation function (ReLU), a 1×1 convolutional layer (conv), and an upsampling operation connected in sequence. For the , Decoding features output by each decoding layer , Since directly performing low-rank sparse decomposition on high-dimensional multi-channel features would incur significant computational overhead, this embodiment first obtains single-channel feature maps by compressing upsampling units (CUBs). , ,in, , and They represent the first Decoding features The number of channels, spatial height, and spatial width, and These represent the input infrared images respectively. Or binary true mask The corresponding spatial height and spatial width. Upsampling uses bilinear interpolation to decode the features. Restored to the binary true mask Matching space dimensions.

[0044] Single-channel feature map When processing the input feature block construction unit, a size of [size missing] is used. Step size is Sliding window for single-channel feature maps Perform local patch block extraction and unfolding, then vectorize each local patch block and stack them column-wise; that is, each local patch block is expanded into a length of... The column vectors are stacked in a sliding spatial order to obtain the feature block matrix. Based on the proof that the features satisfy the IPI prior, , , , ,in, This represents the low-rank background feature matrix. Let N represent the sparse target feature matrix, and let N represent the total number of extracted local patch blocks.

[0045] Single-channel feature map When the input is processed in the Robust Hyperparameter Estimation Unit (RHE), it is sequentially processed by Global Average Pooling (GAP), Multilayer Perceptron (MLP), and Normalized Activation Function (Sigmoid) before outputting iterative decoupling hyperparameters. These iterative decoupling hyperparameters include the target sparse regularization parameters obtained in each iteration during the training phase. and augmented Lagrange penalty parameters , Used to control sparse target components The degree of sparsity constraint, Used to constrain the decomposition residual DBT and adjust the threshold strength and convergence stability during the iterative update process.

[0046] Single-channel feature maps are extracted using Global Average Pooling (GAP) and Multilayer Perceptron (MLP). The global context is used to adaptively predict the first... The target sparse regularization parameter required for the next iteration and augmented Lagrange penalty parameters ,as follows:

[0047] in, , This indicates the total number of iterations.

[0048] A robust hyperparameter estimation unit (RHE) was designed to adapt to different infrared scenarios, different decoding levels, and different sample feature distributions.

[0049] Feature block matrix When the hyperparameters of the iterative decoupling process are input together into the iterative decoupling unit for processing, a constrained optimization problem is established. as follows:

[0050] in, and Representing the feature block matrix respectively The low-rank background feature matrix and sparse target feature matrix; Represents the Lagrange multiplier matrix; Represents the nuclear norm. express Norm, Represents the matrix inner product. This represents the Frobenius norm.

[0051] The iterative decoupling unit will decouple the feature block matrix The decoupling is represented as a nonconvex problem in Robust Principal Component Analysis (RPCA) as follows:

[0052] in, This represents the rank function of a matrix, used to measure the size of the rank of a matrix. Used to constrain the background matrix The low-rank nature; express Norm, Used to constrain the target matrix Sparsity.

[0053] Due to rank ( )and Direct optimization of norms is difficult; the above non-convex problems utilize nuclear norms and Norm relaxation is a convex optimization problem for ease of computation, as follows:

[0054] We also use the augmented Lagrangian function to efficiently solve convex optimization problems.

[0055] Then, the inaccurate augmented Lagrange multiplier method IALM is applied. The next iteration updates explicitly separate the low-rank background feature matrix. Sparse target feature matrix Initial settings , , Number of iterations Set to 5 to strike a balance between background suppression and weak target preservation. In the... In this iteration, the sparse target feature matrix is ​​processed using the Alternating Direction Multiplier Method (ADMM). The following updates have been made:

[0056]

[0057]

[0058] in, Indicates the first The sparse target feature matrix of the next iteration; Indicates based on parameters The soft threshold operator, , and They represent the first The target sparse regularization parameter and the augmented Lagrange penalty parameter for the next iteration. ; Indicates the first The low-rank background feature matrix of the next iteration; and They represent the first Subsequent The Lagrange multiplier matrix of the next iteration; Indicates based on parameters The Singular Value Thresholding (SVT) operator in background feature matrix updates For matrix Its singular value decomposition is Then there is ,in, and Let these represent the left and right singular vector matrices obtained from the singular value decomposition, respectively. Let represent a diagonal matrix composed of singular values. This is a transpose.

[0059] go through After several iterations, the final output is the decoupled sparse target feature matrix. .

[0060] Sparse target feature matrix During the input target reconstruction supervision unit, feature reconstruction (Fold) and overlapping region averaging (AOR) are performed sequentially to reconstruct the spatial domain target feature map. Then, after processing by the target thinning convolutional unit (TRC), an auxiliary prediction mask is output. The target refinement convolutional unit (TRC) consists of a series of 3×3 convolutional layers (conv), an activation function (ReLU), and another 3×3 convolutional layer (conv).

[0061] Specifically, in obtaining the first Decoding layer The sparse target feature matrix after the second iteration Then, first use the Fold operation to... Each column is restored to its corresponding Local target feature blocks, and return them to their original positions. The original spatial location within the sliding window. Since there are overlapping regions between the sliding windows, the overlapping regions are averaged and fused using AOR to obtain a spatial domain target feature map. This is done to avoid blocky artifacts and obtain a smooth target response map. Then, Input target thinning convolutional unit TRC output auxiliary prediction mask TRC is used to enhance the response of sparse targets and refine the target boundaries.

[0062] Total loss function Specifically as follows:

[0063]

[0064] in, and These represent the main network segmentation loss and the IGFD module auxiliary loss, respectively. The main network segmentation loss specifically adopts the Soft-IoU (Soft Intersection over Union Loss) loss. ; This represents the main prediction mask output by the detection head; Represents a binary true mask; α Represents the loss balance coefficient. α We set the value to 0.10 to achieve a balance between IPI prior constraints and main detection task optimization. Indicates Soft-IoU loss; The decoder's first l The auxiliary prediction mask obtained from the layer decoding layer.

[0065] Small infrared targets typically occupy only a very small portion of the entire image, resulting in a severe imbalance between positive and negative samples. Therefore, Soft-IoU loss is used for joint optimization in both the main detection branch and the IGFFD auxiliary branch to ensure consistency in gradient direction. The Soft-IoU loss is expressed as:

[0066] in, This represents the predicted image output by the detection head. This represents element-wise multiplication; Represents the numerically stable term, taking The value is 1.

[0067] This method introduces the low-rank background and sparse target priors from the infrared block image model IPI into the decoding feature space to constrain the decoder to encode stable background-target structure feature representations, thereby internalizing the IPI priors and improving the robustness and generalization performance of the infrared small target detection model in the unknown domain.

[0068] After model training is complete, the IGIFD module is completely removed during the inference phase. Therefore, this invention internalizes the IPI prior into the trained detection network parameters without altering the network structure and computational path during inference. This training-inference separation design enables the model to achieve greater robustness in unknown domains while maintaining inference speed essentially consistent with the original baseline network.

[0069] During training, the Adam optimizer is used with an initial learning rate of The batch size is 8, and the number of training epochs is 200. The input image size is... All models are trained end-to-end, and multiple units in IGFFD participate in backpropagation during the training phase.

[0070] Finally, the infrared image of the target domain to be detected is input into the trained infrared small target detection network. After processing, the detection results of cross-domain small targets in the target domain infrared image are output. In a specific application, infrared small target datasets such as SIRST, SIRSTV2, and IRSTD-1K can be used to construct cross-domain training and testing scenarios. For example, the model is trained only on the source domain dataset and then tested directly on the unseen target domain dataset. The source and target domains can come from different sensors, different background scenes, or different data acquisition conditions. After training is completed, the optimal model weights are loaded for testing. An unseen target image is input, and the infrared small target detection results are obtained through forward inference of the network. Figure 3 As shown, after introducing IGFD into multiple baseline networks, both false negatives and false positives in cross-domain detection results are suppressed, and the response in the target region becomes more concentrated. This demonstrates that the method of this invention significantly improves the generalization ability of the model.

[0071] This invention provides a practical and technologically advanced method for generalizing the infrared small target detection domain, which has good engineering implementation value and broad application prospects.

[0072] The above embodiments are only used to explain the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions or improvements made to the network backbone, number of convolutional layers, number of channels, window size, number of iterations, loss function, or training parameters based on the technical concept of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A cross-domain infrared small target detection method based on IPI-guided feature decoupling, characterized in that, include: Step S1: Acquire several source domain infrared images and label them with binary ground truth masks to use as the training set; Step S2: Construct an IPI-guided feature decoupling module IGFD, which includes a compression upsampling unit, a feature block construction unit, a robust hyperparameter estimation unit, an iterative decoupling unit, and a target reconstruction supervision unit. Introduce the IPI-guided feature decoupling module IGFD into the decoder of the infrared small target detection network, and construct the total loss function of the infrared small target detection network, which includes the main network segmentation loss and the IGFD module auxiliary loss. Input the training set into the infrared small target detection network for training until the total loss function converges. Remove the IPI-guided feature decoupling module IGFD to obtain the trained infrared small target detection network. Step S3: Input the infrared image of the target domain to be detected into the trained infrared small target detection network, process it, and output the detection results of cross-domain small targets in the infrared image of the target domain.

2. The cross-domain infrared small target detection method based on IPI-guided feature decoupling according to claim 1, characterized in that: In step S2, the infrared small target detection network includes an encoder, a decoder, and a detection head connected in sequence. The decoder includes L decoding layers. Each decoding layer receives the output of the encoder, processes it, and outputs its own decoding features. The detection head outputs the main prediction mask. During the training phase of the infrared small target detection network, an IPI-guided feature decoupling module (IGFD) is introduced after each decoding layer, but it is not connected to the next decoding layer. For each decoding layer, the output decoded features are first processed by the compression upsampling unit to output a single-channel feature map. Then, they are respectively input into the feature block construction unit and the robust hyperparameter estimation unit for processing, and the feature block matrix and iterative decoupling hyperparameters are output respectively. Then, they are jointly input into the iterative decoupling unit for processing, and the sparse target feature matrix is ​​output. Then, they are input into the target reconstruction supervision unit for processing, and the auxiliary prediction mask is output. The total loss function is constructed based on the auxiliary prediction mask, the binary true mask, and the main prediction mask.

3. The cross-domain infrared small target detection method based on IPI-guided feature decoupling according to claim 2, characterized in that: In step S2, when the decoded feature input is processed in the compression and upsampling unit CUB, feature compression is performed first, followed by upsampling, and then a single-channel feature map is output. .

4. The cross-domain infrared small target detection method based on IPI-guided feature decoupling according to claim 2, characterized in that: In step S2, the single-channel feature map When processing the input feature block construction unit, a size of [size missing] is used. Step size is Sliding window for single-channel feature maps Local patch blocks are extracted, then each local patch block is vectorized and stacked column-wise to obtain the feature block matrix. , ,in, This represents the low-rank background feature matrix. This represents the sparse target feature matrix.

5. The cross-domain infrared small target detection method based on IPI-guided feature decoupling according to claim 2, characterized in that: In step S2, the single-channel feature map When the input is processed in the Robust Hyperparameter Estimation Unit (RHE), it is sequentially processed by Global Average Pooling (GAP), Multilayer Perceptron (MLP), and Normalized Activation Function (Sigmoid) before outputting iterative decoupling hyperparameters. These iterative decoupling hyperparameters include the target sparse regularization parameters obtained in each iteration during the training phase. and augmented Lagrange penalty parameters .

6. The cross-domain infrared small target detection method based on IPI-guided feature decoupling according to claim 5, characterized in that: In step S2, the feature block matrix When the hyperparameters of the iterative decoupling process are input together into the iterative decoupling unit for processing, a constrained optimization problem is established. as follows: in, and Representing the feature block matrix respectively The low-rank background feature matrix and sparse target feature matrix; Represents the Lagrange multiplier matrix; Represents the nuclear norm. express Norm, Represents the inner product of matrices. Denotes the Frobenius norm; Then, the inaccurate augmented Lagrange multiplier method IALM is applied. The next iteration updates explicitly separate the low-rank background feature matrix. Sparse target feature matrix In the In this iteration, the sparse target feature matrix is ​​processed using the Alternating Direction Multiplier Method (ADMM). The following updates have been made: in, Indicates the first The sparse target feature matrix of the next iteration; Indicates based on parameters The soft threshold operator, , and They represent the first The target sparse regularization parameter and the augmented Lagrange penalty parameter for the next iteration; Indicates the first The low-rank background feature matrix of the next iteration; and They represent the first Subsequent The Lagrange multiplier matrix of the next iteration; Indicates based on parameters The singular value threshold (SVT) operator, ; go through After several iterations, the final output is the decoupled sparse target feature matrix. .

7. The cross-domain infrared small target detection method based on IPI-guided feature decoupling according to claim 2, characterized in that: In step S2, the sparse target feature matrix During the input target reconstruction supervision unit, feature reconstruction (Fold) and overlapping region averaging (AOR) are performed sequentially to reconstruct the spatial domain target feature map. Then, after processing by the target thinning convolutional unit (TRC), an auxiliary prediction mask is output. .

8. The cross-domain infrared small target detection method based on IPI-guided feature decoupling according to claim 2, characterized in that: In step S2, the total loss function Specifically as follows: in, and These represent the main network segmentation loss and the IGFFD module auxiliary loss, respectively. This represents the main prediction mask output by the detection head; Represents a binary true mask; α Indicates the loss balance coefficient; Indicates Soft-IoU loss; The decoder's first l The auxiliary prediction mask obtained from the layer decoding layer.

9. An electronic device, characterized in that, include: A memory and a processor are coupled to each other, wherein the memory stores program data, and the processor invokes the program data to perform the method as described in any one of claims 1-8.

10. A computer-readable storage medium storing program data thereon, characterized in that, When the program data is executed by the processor, it implements the method as described in any one of claims 1-8.

Images