An infrared small target detection method and system based on frequency-guided one-step denoising

Abstract

The application discloses an infrared small target detection method and system based on frequency-guided one-step denoising, and the method comprises the following steps: constructing a diffusion framework composed of a reconstruction network and a detection network and training the diffusion framework; the reconstruction network adopts a denoising diffusion probability model to perform one-step denoising on an input infrared image, directly estimates a target background from an image containing noise disturbance to ensure inference efficiency, introduces a frequency-guided perception loss, and effectively improves the reconstruction quality of the background image in terms of details and textures by constraining the high-frequency components of the reconstructed background and the real background; finally, the original infrared image and the reconstructed target-free background image are spliced along the channel dimension to form a two-channel input, and the two-channel input is sent into the detection network to learn the difference between the two channels and accurately predict a target mask; under the premise of ensuring real-time inference efficiency, the application fully utilizes the distribution modeling capability of the generative model and accurately reconstructs the background image containing discriminative high-frequency information.

Description

Technical Field

[0001] This invention relates to the field of infrared image detection and processing technology, specifically to a method for detecting small infrared targets based on frequency-guided one-step denoising.

[0002] The present invention also relates to an infrared small target detection system based on frequency-guided one-step denoising. Background Technology

[0003] Infrared Small Target Detection (IRSTD) aims to accurately identify and locate targets that are extremely small in size, have low signal-to-noise ratio, and lack structural and textural features from infrared images. Due to the long imaging distance, high sensor noise, and strong background clutter in infrared imaging, targets are often submerged in noise and background, making it impossible to balance robustness and accuracy in the detection process.

[0004] Current technological developments are as follows: With the development of deep learning, the performance of IRSTD has been significantly improved. Early work was mainly based on the U-Net architecture, which enhanced the spatial response of small targets through skip connections and multi-scale feature fusion. Subsequently, the Transformer model was introduced with a self-attention mechanism to capture global long-range dependencies, further improving detection accuracy. Recently, visual Mamba and the large segmentation model SAM (Segment Anything Model) have been explored to enrich semantic representations. However, the above methods all essentially follow a discriminative paradigm—that is, learning a deterministic mapping from infrared images to target masks. This paradigm inherently relies on a high degree of fit to the distribution of training data. When faced with out-of-distribution backgrounds or extreme noise, its generalization ability drops sharply, and the output lacks the ability to model the uncertainty of the generated distribution.

[0005] With technological advancements, diffusion models have demonstrated powerful generative capabilities in image synthesis and restoration in recent years. Existing diffusion model-based methods primarily follow two paths: one utilizes diffusion models to generate additional training samples to alleviate data scarcity, i.e., using diffusion models to synthesize labeled infrared small target samples to expand the training set. However, infrared small targets correspond only to sparse peaks at extremely high frequencies in the frequency domain, and their boundaries transition smoothly with background pixels. During backsampling, the lack of high-frequency constraints makes it easy for diffusion models to treat target signals as noise and erase them, or generate false targets with blurred boundaries and distorted intensity. Such synthesized data not only fails to provide effective supervisory signals but also introduces... The first approach involves the distribution offset and the discrimination boundary of the contamination feature extractor. The second approach uses a diffusion network as an iterative decoder to gradually refine the predicted mask, that is, to gradually convert random noise into target mask prediction through a multi-step denoising process. Although this paradigm utilizes the distribution fitting advantage of the generative model, it relies on multi-step iterative sampling to approximate the real mask distribution. In the real-time processing scenario of IRSTD, this serial computation mode introduces high inference latency and memory usage. More importantly, this process directly applies the denoising conditions to the mask space, completely ignoring the decisive role of the accuracy of background reconstruction in target localization. Clutter residue caused by background estimation bias will directly mask weak target signals in the pixel-by-pixel loss.

[0006] In summary, for infrared small target detection, high-frequency components in the background are often the key spatial clues for distinguishing targets from false alarms. Existing loss functions cannot constrain the model network to reconstruct high-frequency components in the background with high fidelity, resulting in an overly smooth reconstructed background that loses key discriminative information.

[0007] The problem with existing technologies is: how to fully utilize the distributed modeling capabilities of generative models and accurately reconstruct background images containing discriminative high-frequency information while ensuring real-time inference efficiency.

[0008] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] This invention provides an infrared small target detection method and system based on frequency-guided one-step denoising, which can effectively overcome the defects existing in the prior art.

[0010] A frequency-guided one-step denoising method for infrared small target detection, the method comprising: The infrared image of the target to be tested is input into the pre-trained reconstruction-detection model for background reconstruction and mask generation to obtain the segmentation mask of the target to be tested. The reconstruction-detection model includes a reconstruction network and a detection network, and the training process includes: Acquire an infrared small target image as the original image, and then obtain a standardized infrared small target image based on the original image; The reconstruction network is a denoising diffusion probability model. It receives the standardized infrared small target image and performs Gaussian noise addition processing to obtain a noisy image. The noisy image is then denoised to obtain the reconstructed background image. The detection network is a U-Net architecture with a channel attention module. It receives the original image and the reconstructed background image, performs channel stitching on the original image and the reconstructed background image along the channel dimension, performs feature map extraction after the channel stitching process, and outputs a segmentation mask for infrared small targets. In the training process of the reconstruction network, a frequency-guided perceptual loss is constructed based on the background image and the ground truth of the background, and the reconstruction network is optimized by the frequency-guided perceptual loss.

[0011] In some exemplary embodiments, after the reconstruction network receives the standardized infrared small target image, the method further includes: treating the small targets in the standardized infrared small target image as noise perturbations, adding Gaussian noise to the input standardized infrared small target image to a preset time step using forward diffusion processing to obtain a noisy image, and denoising the noisy image using a one-step denoising method to obtain a target-free background image.

[0012] In some exemplary embodiments, the method further includes, during the training process of reconstructing the network: The total training loss of the reconstruction network is a weighted combination of the baseline mean square error loss in the spatial domain between the reconstructed background image and the ground truth background, and the frequency-guided perceptual loss. The construction of the frequency-guided perception loss includes: performing two-dimensional discrete Fourier transform on the acquired background image and the true background value to the frequency domain, and then constructing a high-pass filter mask to retain the high-frequency components of the background image and the true background value transformed to the frequency domain. The mean square error between the high-frequency components of the background image and the true background value in the frequency domain is used as the frequency-guided perception loss. The frequency-guided perception loss is used to constrain the consistency between the reconstructed background image and the true background value in the high-frequency domain.

[0013] In some exemplary embodiments, the method further includes, during the training process of reconstructing the network: Morphological operations are used to process the original image to generate targetless background images as pseudo-labels, which are then used to supervise the training of the reconstruction network.

[0014] In some exemplary embodiments, after the detection network receives the original image and the reconstructed background image, the method further includes: The original image and the reconstructed background image are concatenated along the channel dimension to form a dual-channel input. The encoder extracts multi-scale semantic feature maps of the original image and the reconstructed background image. The channel attention module rescales the multi-scale semantic feature map, and finally the rescaled multi-scale semantic feature map is processed by the decoder to restore spatial resolution and output a segmentation mask for infrared small targets.

[0015] In some exemplary embodiments, the method further includes, during the training of the detection network: After channel-wise concatenation of the original image and the reconstructed background image, the image is input into the detection network for end-to-end training. The total training loss of the detection network is constructed by weighting the binary cross-entropy loss and the soft-IoU loss during training.

[0016] In some exemplary embodiments, during the training of the detection network, the method further includes: processing the backgroundless image without a target output by the reconstruction network through gradient blocking, and then performing channel concatenation with the original image and inputting it into the detection network for training.

[0017] According to a second aspect of the present invention, an infrared small target detection system based on frequency-guided one-step denoising is provided, the system comprising: The data acquisition module is used to acquire infrared small target images as raw images and to obtain standardized infrared small target images based on the raw images; The reconstruction-detection module is used to process standardized infrared small target images to obtain segmentation masks for infrared small targets; the target detection module includes a reconstruction network module and a detection network module. The reconstruction network module is used to add Gaussian noise to the standardized infrared small target image to obtain a noisy image, and then denoise the noisy image to obtain a reconstructed background image. The detection network module is used to receive the original image and the reconstructed background image, perform channel stitching of the original image and the reconstructed background image along the channel dimension, perform feature map extraction processing after the channel stitching process is completed, and output the segmentation mask of infrared small target. The result output module is used to perform resolution consistency processing on the segmentation mask and output the final target detection result.

[0018] In some exemplary embodiments, the detection network is a U-Net architecture with a channel attention module, including: an encoder, a decoder, and skip connections; The encoder is used to extract multi-scale semantic feature maps of the original image and the reconstructed background image through step-by-step downsampling; the skip connection is used to fuse features from the corresponding encoder and decoder stages to preserve fine-grained spatial details; the channel attention module is embedded in each coding block of the encoder and is used to perform global max pooling and global average pooling on the input multi-scale semantic feature map, and then rescale the multi-scale semantic feature map after calculating the channel weights through a multilayer perceptron and a sigmoid activation function; the decoder is used to restore the spatial resolution of the scaled multi-scale semantic feature map through step-by-step upsampling.

[0019] In some exemplary embodiments, the reconstruction network module and the detection network module in the target detection module are decoupled, so that the reconstruction network module and the detection network module can be independently configured.

[0020] Due to the adoption of the above technical solution, the beneficial effects of the present invention are as follows: The infrared small target detection method of this invention is based on background image reconstruction. It treats infrared small targets as noise perturbations and uses the denoising characteristics of the diffusion model to recover the target-free background image from the perturbation image, thus completing the reconstruction of the background image. In the process of background image reconstruction, frequency-guided perceptual loss is introduced. By performing high-pass filtering on the reconstructed background and the ground background in the frequency domain and calculating the mean square error of the high-frequency components, the reconstruction consistency in the high-frequency domain is constrained, thereby effectively improving the reconstruction quality of the background image in terms of details and texture and avoiding false high-frequency artifacts introduced by reconstruction errors. At the same time, a one-step denoising strategy is adopted to skip multiple iterative sampling steps, which significantly improves inference efficiency while ensuring reconstruction quality.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0023] Figure 1 A schematic diagram illustrating the process of the method of the present invention is shown. Figure 2 This schematic diagram illustrates the process of forward diffusion noise addition and denoising in the reconstruction network of the method of the present invention; Figure 3 The diagram illustrates the architecture of the detection network in the method of the present invention. Figure 4 The schematic diagram illustrates the process of introducing frequency-guided sensing loss into the reconstruction network in the method of the present invention. Detailed Implementation

[0024] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0025] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0026] Existing infrared small target detection methods lack a way to fully utilize generative modeling capabilities and perform infrared small target detection in an efficient reasoning manner, so as to achieve effective decoupling between background reconstruction and target detection, and simultaneously ensure detection accuracy and reasoning efficiency under complex background and low signal-to-noise ratio conditions.

[0027] To address the shortcomings and deficiencies of existing technologies, this example implementation proposes a frequency-guided one-step denoising method for infrared small target detection. Infrared small targets are treated as noise-like perturbations. The denoising characteristics of the reconstruction network based on a denoising diffusion probability model are used to recover the target-free background image from the infrared image of the target containing the perturbation. The residual between the original image and the reconstructed background image is calculated to highlight candidate target regions, providing explicit priors for subsequent detection. Frequency-guided perceptual loss is introduced during reconstruction. By performing high-pass filtering on the reconstructed background and the ground truth background in the frequency domain and calculating the mean square error of the high-frequency components, the reconstruction consistency in the high-frequency domain is constrained, effectively improving the reconstruction quality of the background image in terms of detail and texture. Simultaneously, a one-step denoising strategy is adopted to skip multiple iterative sampling steps, significantly improving inference efficiency while ensuring reconstruction quality. The detection network uses the dual-channel representation formed by stitching the original image and the reconstructed background image as input to the network layers, accurately predicting the target segmentation mask by learning the difference between the two channels. The reconstruction network of this application is a plug-and-play module, compatible with existing detection models, effectively improving detection accuracy and efficiency without increasing the computational complexity of the detection network.

[0028] In this example implementation, the hardware environment can be a server with 4 NVIDIA RTX 4090D cards (24GB), the operating system can be Ubuntu 20.04, and the Python version is 3.8. The experiment uses three public infrared small target detection datasets: NUAA-SIRST, NUDT-SIRST, and IRSTD-1K, with a uniform resolution of 256×256, and the training / test ratio is 5:5 or 8:2.

[0029] refer to Figure 1 As shown, this example provides a frequency-guided one-step denoising method for infrared small target detection. It includes constructing an infrared small target detection framework based on a diffusion model, i.e., a reconstruction-detection model, for training. Then, the infrared image of the target to be detected is input into the trained reconstruction-detection model for background reconstruction and mask generation to obtain a segmentation mask for the target. The method includes the following steps: For example, the training process of the reconstruction-detection model described above specifically includes: In step S1, an infrared small target image is acquired as the original image, and a standardized infrared small target image is acquired based on the original image.

[0030] Specifically, in this example, the reconstruction-detection model can be trained by selecting one of the three publicly available infrared small target detection datasets, NUAA-SIRST, NUDT-SIRST, and IRSTD-1K, and the original image can be standardized to obtain a standardized infrared small target image.

[0031] In step S2, a reconstruction-detection model is constructed, which includes a reconstruction network and a detection network.

[0032] For example, the reconstruction network is a denoising diffusion probability model, which receives the standardized infrared small target image and performs Gaussian noise addition processing to obtain a noisy image, and then performs denoising processing on the noisy image to obtain a reconstructed background image.

[0033] For example, the reconstruction network adopts a denoising diffusion probability model, the core idea of ​​which is to treat small targets in infrared images as noise-like perturbations. From the perspective of infrared imaging, small infrared targets usually appear as local "anomalies" with intensity differences from their surrounding background. However, due to factors such as blurred target boundaries, complex background textures, and low signal-to-noise ratio, these "anomalies" are often masked. Based on this analysis, this example models small infrared targets as noise-like perturbations superimposed on a smooth background, and uses the denoising diffusion probability model to suppress or remove foreground targets, thereby reconstructing a background image without targets.

[0034] Specifically, refer to Figure 2 As shown, given an input infrared small target image During the forward diffusion process, according to the set time steps Gaussian noise is gradually added to the image to obtain a noisy image. The formula for calculating the forward diffusion process is: (1) In the formula, Standard Gaussian noise, To keep pace with time The corresponding noise scheduling parameters; as noise is gradually added, the target signal is submerged by noise, and the noisy image gradually tends to a pure Gaussian noise distribution.

[0035] Then, the noisy image Input the reconstruction network and use a one-step denoising strategy to directly estimate the background image without the target. Skip the multi-step iterative sampling process, as shown in the following equation: (2) In the formula, To reconstruct the noise components predicted by the network, a one-step denoising strategy significantly accelerates the inference process and reduces computational cost by skipping intermediate iterations; time steps The selection of [aspect name] is crucial to the quality of reconstruction. When the noise level is too high, excessive noise can easily cause distortion and artifacts in the first-step denoising process, blurring background details. When the step size is too small, the target signal, which resembles noise, is difficult to remove in one step. This example experiment shows that within the step range of 30 to 90 steps, the overall performance change is relatively small, and for medium step sizes such as... or Preferring to produce better results, this example implementation is preferred. .

[0036] For example, frequency-guided perceptual loss is introduced into the training of the reconstruction network. The reconstructed background image and the ground truth background are subjected to two-dimensional discrete Fourier transforms respectively. High-frequency components are preserved by high-pass filtering masks. The mean square error between the high-frequency representations of the two is calculated as the frequency-guided perceptual loss, which constrains the consistency of the reconstructed background in the high-frequency domain and improves the reconstruction quality of the background image in terms of details and texture. The mean square error loss between the reconstructed background and the ground truth background and the frequency-guided perceptual loss are weighted and combined as the total training loss of the reconstruction network. Infrared small targets usually occupy only a small part of the image area, which makes the detection accuracy extremely sensitive to the background reconstruction quality. Even small background reconstruction errors may introduce false high-frequency artifacts, interfering with the judgment of subsequent detection networks. These high-frequency components, especially edges and textures, provide important clues for distinguishing targets from complex background clutter.

[0037] Specifically, to explicitly preserve these structural details, this example introduces high-frequency sensing constraints in the frequency domain; see reference. Figure 4 As shown, the background image will be reconstructed. With background truth value Two-dimensional discrete Fourier transforms are performed to shift the zero-frequency component to the center of the spectrum, and the logarithmic magnitude spectrum is calculated and spatially normalized. Then, a binary high-pass filter mask is constructed. The radial distance in the retained spectrum exceeds the cutoff frequency. The high-frequency components; where, for the radial distance in the spectrum, satisfying The frequency component is assigned a value of 1, and the rest are assigned a value of 0. The coordinates of the spectrum center; high frequency representation The frequency-guided perceptual loss is obtained by element-wise multiplication of the high-pass filter mask and the normalized logarithmic magnitude spectrum. It is defined as the mean square error between the high-frequency spectrum of the reconstructed background and the true background value. This frequency-guided perceptual loss prompts the reconstruction network to pay more attention to the high-frequency reconstruction pressure introduced by the target region, while maintaining the intrinsic structure of the background in the one-step denoising process, thereby generating a higher quality target-free background image.

[0038] For example, the detection network is a U-Net architecture with a channel attention module. It receives the original image and the reconstructed background image, performs channel stitching of the original image and the reconstructed background image along the channel dimension, performs feature map extraction processing after the channel stitching process is completed, and outputs a segmentation mask for infrared small targets.

[0039] For example, refer to Figure 3 As shown, the detection network adopts the standard U-Net architecture with a channel attention module, including an encoder, a decoder, and skip connections; first, the original input infrared small target image is processed... With reconstructed background image without target By splicing along the channel dimension, a dual-channel input is formed. This dual-channel representation enables the detection network to perceive the target region by learning the difference between the original image and the reconstructed background image.

[0040] Specifically, the encoder of the detection network extracts multi-scale semantic features through stepwise downsampling, while the decoder gradually restores spatial resolution through stepwise upsampling. Skip connections fuse features from the encoder and decoder stages, helping to preserve fine-grained spatial details crucial for small target localization. A channel attention module is inserted after each encoding block of the encoder, performing global max pooling and global average pooling on the input feature map. The sum of these two modules is then used to calculate channel weights via a multilayer perceptron and a sigmoid activation function, rescaling the feature map by channel dimension. This module enables the network to adaptively emphasize channel features more relevant to small targets while suppressing background interference. This allows for a clearer evaluation of the proposed reconstruction network's effectiveness and directly verifies the core motivation of this example. Furthermore, unlike previous diffusion-based methods, the reconstruction and detection networks in this example are decoupled, providing greater flexibility in practical deployment: depending on computational resource constraints, the detection network can be replaced with a lighter or more powerful model. Importantly, the plug-and-play nature of this example allows for direct use with existing infrared small target detection models. Experimental results show that integrating the reconstruction network with various advanced detectors improves their performance.

[0041] In step S3, the preprocessed infrared small target image is input into the reconstruction-detection model framework constructed in step S2, and the network is trained in an end-to-end manner to obtain the trained reconstruction-detection model. During the training phase, morphological operations are first applied to the original image to generate a target-free background image as a pseudo-label to supervise the training of the reconstruction network. The reconstruction network and the detection network are optimized separately, and the gradient between them is blocked during training. Specifically, the training loss of the reconstruction network is: (3) In the formula, The mean squared error loss between the reconstructed background image and the ground truth in the image domain is used to constrain the overall fidelity of the reconstructed background image. The aforementioned frequency-guided sensing loss is used to constrain the reconstruction consistency in the high-frequency domain; To balance the coefficients; the backgroundless image output by the reconstruction network is subjected to gradient blocking, and then stitched with the original infrared image to form a dual-channel input which is fed into the detection network, ensuring that the gradients of the reconstruction network and the detection network do not interfere with each other and are optimized independently; the training loss of the detection network is: (4) In the formula, For binary cross-entropy loss, For Soft-IoU loss, , These are the corresponding balance coefficients.

[0042] In step S4, the testing process includes: after preprocessing the infrared image of the target to be tested, it is input into the trained reconstruction-detection model. First, the reconstruction network performs forward diffusion noise addition on the input infrared image of the target to be tested up to the time step. Then, the background image without the target is directly estimated through a one-step denoising strategy; subsequently, the original image and the reconstructed background image are concatenated into a dual-channel input and fed into the detection network, which outputs the segmentation mask of the target to be tested; the whole process only requires a single-step denoising to complete the background reconstruction, without the need for multi-step iterative sampling, thus ensuring a high-efficiency inference speed.

[0043] The method provided in this invention significantly improves inference efficiency while ensuring reconstruction quality by employing a noise-like modeling, one-step denoising, and frequency-guided generative reconstruction denoising strategy that skips multiple iterative sampling steps. Frequency-guided perceptual loss effectively improves the reconstruction quality of the background image in terms of detail and texture by constraining reconstruction consistency in the high-frequency domain, avoiding false high-frequency artifacts introduced by reconstruction errors. The dual-channel input design enables the detection network to directly learn the differences between the original image and the reconstructed background, accurately perceiving the target region. The decoupling design of the reconstruction and detection networks allows the reconstruction module to be flexibly integrated with different detection architectures as a plug-and-play component, effectively improving detection performance without increasing the computational complexity of the detection network.

[0044] Experimental results show that the present invention achieves state-of-the-art performance on three public benchmark datasets: IRSTD-1K, NUDT-SIRST, and NUAA-SIRST. Specifically, on the IRSTD-1K dataset, the present invention improves the IoU metric by approximately 5 percentage points compared to the U-Net baseline, and also reduces the false alarm rate. The performance was reduced by approximately 13 percentage points. After integrating the reconstruction network of this invention with different detectors such as ACM, DNANet, and SCTranNet, each detector achieved a consistent performance improvement, fully verifying the generalization ability and practical value of this invention as a plug-and-play module. This invention has the advantages of high detection accuracy, good inference efficiency, and strong deployment flexibility, making it very suitable for infrared small target detection tasks with stringent requirements for real-time performance and accuracy, such as early warning, remote monitoring, and autonomous sensing.

[0045] Furthermore, this example embodiment also provides an infrared small target detection system based on frequency-guided one-step denoising, the system comprising: The data acquisition module is used to acquire infrared small target images as raw images and to obtain standardized infrared small target images based on the raw images; The reconstruction-detection module is used to process standardized infrared small target images to obtain segmentation masks for infrared small targets; the target detection module includes a reconstruction network module and a detection network module. The reconstruction network module is used to add Gaussian noise to the standardized infrared small target image to obtain a noisy image, and then denoise the noisy image to obtain a reconstructed background image. The detection network module is used to receive the original image and the reconstructed background image, perform channel stitching of the original image and the reconstructed background image along the channel dimension, perform feature map extraction processing after the channel stitching process is completed, and output the segmentation mask of infrared small target. The result output module is used to perform resolution consistency processing on the segmentation mask and output the final target detection result.

[0046] The module units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor; wherein, the names of these units do not constitute a limitation on the unit itself under certain circumstances.

[0047] It should be noted that although several modules or units of the device for performing actions are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit; conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be embodied.

[0048] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0049] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention 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 the invention are indicated by the claims.

[0050] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for detecting small infrared targets based on frequency-guided one-step denoising, characterized in that, The method includes: The infrared image of the target to be tested is input into the pre-trained reconstruction-detection model for background reconstruction and mask generation to obtain the segmentation mask of the target to be tested. The reconstruction-detection model includes a reconstruction network and a detection network, and the training process includes: Acquire an infrared small target image as the original image, and then obtain a standardized infrared small target image based on the original image; The reconstruction network is a denoising diffusion probability model. It receives the standardized infrared small target image and performs Gaussian noise addition processing to obtain a noisy image. The noisy image is then denoised to obtain the reconstructed background image. The detection network is a U-Net architecture with a channel attention module. It receives the original image and the reconstructed background image, performs channel stitching on the original image and the reconstructed background image along the channel dimension, performs feature map extraction after the channel stitching process, and outputs a segmentation mask for infrared small targets. In the training process of the reconstruction network, a frequency-guided perceptual loss is constructed based on the background image and the ground truth of the background, and the reconstruction network is optimized by the frequency-guided perceptual loss.

2. The infrared small target detection method based on frequency-guided one-step denoising according to claim 1, characterized in that, After the reconstruction network receives the standardized infrared small target image, the method further includes: treating the small targets in the standardized infrared small target image as noise perturbations, adding Gaussian noise to the input standardized infrared small target image to a preset time step using forward diffusion processing to obtain a noisy image, and performing denoising processing on the noisy image using a one-step denoising method to obtain a target-free background image.

3. The infrared small target detection method based on frequency-guided one-step denoising according to claim 1, characterized in that, During the training process of reconstructing the network, the method further includes: The benchmark mean square error loss in the spatial domain between the reconstructed background image and the ground truth background is weighted and combined with the frequency-guided perceptual loss as the total training loss of the reconstruction network. The construction of the frequency-guided perception loss includes: performing two-dimensional discrete Fourier transform on the acquired background image and the true background value to the frequency domain, and then constructing a high-pass filter mask to retain the high-frequency components of the background image and the true background value transformed to the frequency domain. The mean square error between the high-frequency components of the background image and the true background value in the frequency domain is used as the frequency-guided perception loss. The frequency-guided perception loss is used to constrain the consistency between the reconstructed background image and the true background value in the high-frequency domain.

4. The infrared small target detection method based on frequency-guided one-step denoising according to claim 3, characterized in that, During the training process of reconstructing the network, the method further includes: Morphological operations are used to process the original image to generate targetless background images as pseudo-labels, which are then used to supervise the training of the reconstruction network.

5. The infrared small target detection method based on frequency-guided one-step denoising according to claim 1, characterized in that, After the detection network receives the original image and the reconstructed background image, the method further includes: The original image and the reconstructed background image are concatenated along the channel dimension to form a dual-channel input. The encoder extracts multi-scale semantic feature maps of the original image and the reconstructed background image. The channel attention module rescales the multi-scale semantic feature map, and finally the rescaled multi-scale semantic feature map is processed by the decoder to restore spatial resolution and output a segmentation mask for infrared small targets.

6. The infrared small target detection method based on frequency-guided one-step denoising according to claim 5, characterized in that, During the training of the detection network, the method further includes: After channel-wise concatenation of the original image and the reconstructed background image, the image is input into the detection network for end-to-end training. The total training loss of the detection network is constructed by weighting the binary cross-entropy loss and the soft-IoU loss during training.

7. The infrared small target detection method based on frequency-guided one-step denoising according to claim 6, characterized in that, During the training process of the detection network, the method further includes: processing the background image without a target output by the reconstruction network through gradient blocking, and then concatenating it with the original image through channel stitching before inputting it into the detection network for training.

8. An infrared small target detection system based on frequency-guided one-step denoising, characterized in that, The system includes: The data acquisition module is used to acquire infrared small target images as raw images and to obtain standardized infrared small target images based on the raw images; The reconstruction-detection module is used to process standardized infrared small target images to obtain segmentation masks for infrared small targets; the target detection module includes a reconstruction network module and a detection network module. The reconstruction network module is used to add Gaussian noise to the standardized infrared small target image to obtain a noisy image, and then denoise the noisy image to obtain a reconstructed background image. The detection network module is used to receive the original image and the reconstructed background image, perform channel stitching of the original image and the reconstructed background image along the channel dimension, perform feature map extraction processing after the channel stitching process is completed, and output the segmentation mask of infrared small target. The result output module is used to perform resolution consistency processing on the segmentation mask and output the final target detection result.

9. The infrared small target detection system based on frequency-guided one-step denoising according to claim 8, characterized in that, The detection network is a U-Net architecture with a channel attention module, including: an encoder, a decoder, and skip connections; The encoder is used to extract multi-scale semantic feature maps of the original image and the reconstructed background image through step-by-step downsampling; the skip connection is used to fuse features from the corresponding encoder and decoder stages to preserve fine-grained spatial details; the channel attention module is embedded in each coding block of the encoder and is used to perform global max pooling and global average pooling on the input multi-scale semantic feature map, and then rescale the multi-scale semantic feature map after calculating the channel weights through a multilayer perceptron and a sigmoid activation function; the decoder is used to restore the spatial resolution of the scaled multi-scale semantic feature map through step-by-step upsampling.

10. The infrared small target detection system based on frequency-guided one-step denoising according to claim 8, characterized in that, The target detection module is constructed in a decoupled manner from the reconstruction network module and the detection network module, allowing for independent reconstruction and detection of the network modules.

Images