An infrared small target detection method based on dynamic prior and bidirectional interaction fusion

Abstract

The present application belongs to the technical field of infrared image target detection, and particularly relates to an infrared small target detection method based on dynamic prior and bidirectional interactive fusion. In view of the problems of small target scale, low signal-to-noise ratio, easy feature attenuation, static fixed prior information, lack of deep feature collaborative optimization and the like in the existing infrared small target detection method, a dynamic prior driven infrared small target detection method is proposed. The method firstly performs feature extraction on the input infrared image, constructs a dynamic multi-scale local contrast prior generation module, and adaptively establishes target saliency prior information; then a bidirectional interactive fusion mechanism between the prior and the deep feature is designed, the target response is enhanced by using the prior, the prior is dynamically corrected by the deep semantic information at the same time, and the collaborative optimization of the prior and the feature is realized; further, a structure prior distillation constraint is introduced, the target region expression ability is enhanced and the complex background interference is suppressed from the spatial response and structure consistency angles; finally, the target detection result is output through a decoding prediction module. The present application can effectively improve the detection precision and robustness of the infrared weak small target in the complex background, improve the target saliency response and background suppression ability, and is suitable for application scenarios such as low-altitude security, unmanned aerial vehicle monitoring, infrared early warning and target perception.

Description

Technical Field

[0001] This invention belongs to the fields of infrared image processing and deep learning for small target detection, and specifically relates to an infrared small target detection method based on dynamic prior and bidirectional interactive fusion. Background Technology

[0002] With the continuous advancement of my country's low-altitude economic strategy, the number of civilian consumer drones and industrial inspection drones has experienced explosive growth, and drones are being deployed on a large scale in fields such as aerial photography and mapping, logistics and distribution, power line inspection, and environmental monitoring. Compared with traditional large flying targets, consumer micro drones are small in size, fly at low altitudes, are highly maneuverable and stealthy, making them difficult to effectively control using conventional monitoring methods.

[0003] In complex monitoring scenarios at night, during dawn and dusk, and at long distances, visible light imaging equipment is severely affected by light degradation and environmental obstruction, making it unable to effectively capture targets. In contrast, infrared imaging equipment relies on target thermal radiation imaging, is not limited by lighting conditions, and is the core equipment for all-weather low-altitude target monitoring. However, in actual engineering scenarios, long-range drones occupy only a few to tens of pixels in infrared images, making them typical small infrared targets with inherent characteristics such as sparse scale, lack of texture, extremely low signal-to-noise ratio, and blurred target edges. At the same time, natural scenes contain complex background interference such as cloud clutter, building heat sources, high-temperature areas on the ground, and sky gradient noise, which can easily obscure the characteristics of small drone targets, leading to problems such as missed target detection and false background detection, greatly limiting the reliability of infrared detection systems.

[0004] Currently, infrared small target detection technologies at home and abroad are mainly divided into two categories: traditional manual feature detection algorithms and deep learning intelligent detection algorithms. Both types of technologies have inherent defects that are difficult to overcome, and cannot meet the detection requirements of high precision, low false alarm and real-time in complex low-altitude scenarios.

[0005] Traditional infrared small target detection algorithms are based on manually designed feature operators. Typical algorithms include Top-Hat morphological filtering, multi-scale local contrast filtering (WSLCM), maximum median filtering, and image patch sparse representation (IPI). The core principle of these algorithms is to extract salient features of the target by relying on manually designed fixed feature rules, fixed convolution windows, and fixed enhancement and suppression parameters. Its advantages include fast inference speed, no training required, and simple structure, but its shortcomings are extremely prominent: First, all parameters are preset fixed values, which cannot be adaptively adjusted according to the background complexity, noise intensity, and target scale of real-time images, resulting in extremely poor adaptability to different scenarios such as smooth sky backgrounds, complex building clutter, and sea surface gradient heat sources; Second, the fixed window scale cannot adapt to multi-scale small UAV targets, and is prone to problems such as missed detection of small targets and over-suppression of large targets; Third, in the context of strong clutter, manual operators are prone to misjudging bright background noise as targets, resulting in a high false alarm rate; Fourth, traditional algorithms have weak feature expression capabilities and cannot mine deep semantic features, resulting in a precipitous drop in detection performance in extremely low signal-to-noise ratio scenarios, making them unsuitable for complex low-altitude security scenarios.

[0006] To overcome the shortcomings of traditional manual algorithms, deep learning-based infrared small target detection algorithms have become a mainstream research direction. Semantic segmentation networks based on the U-Net encoder-decoder architecture, with their excellent detail feature preservation capabilities, have become the backbone network for infrared small target detection. Existing improved deep learning algorithms mainly fall into three categories: feature enhancement, prior guidance, and multi-scale fusion. Among them, prior guidance algorithms introduce manual priors such as local contrast and Gaussian saliency to assist the network in strengthening small target features, significantly improving detection accuracy in complex backgrounds. Representative algorithms include ALCNet and UIU-Net.

[0007] However, existing deep learning prior detection schemes still have significant technical bottlenecks and cannot meet the needs of engineering implementation: (1) Existing priors are all static fixed priors, and the prior generation parameters are globally fixed. They cannot be dynamically and adaptively adjusted according to the background distribution and target features of a single image. In non-uniform infrared backgrounds and strong and weak clutter scenes, static priors are prone to amplifying background noise and generating a large number of pseudo-target responses; (2) Existing prior fusion modes are all one-way injection mechanisms. They rely only on priors to guide network feature enhancement and cannot use deep semantic features of the network to reverse correct and optimize prior defects. Errors in the initial priors will continue to accumulate in the forward propagation of the network, eventually leading to detection result deviation and contour fragmentation; (3) Existing networks lack structured constraint mechanisms and rely only on pixel-level loss to optimize the model. They cannot constrain the target edge structure and global feature consistency. Small target detection results have problems such as edge defects, positioning deviation, and target breakage; (4) Existing improved network structures are redundant, have a large number of parameters, high computational complexity, and slow inference speed. They cannot be adapted to devices with low computing power and high real-time requirements such as infrared PTZ and embedded terminals, resulting in poor engineering implementation.

[0008] In summary, current infrared small target detection technology suffers from core technical problems such as poor static prior adaptive capability, accumulation of one-way fusion errors, insufficient structural consistency, and the inability to balance real-time performance and accuracy. There is an urgent need for a novel infrared small target detection solution that can adapt to complex scenarios, achieve bidirectional collaborative optimization, and is lightweight and highly accurate, in order to solve the problem of detecting low-altitude micro-UAVs. Summary of the Invention

[0009] To address the aforementioned problems, the present invention aims to provide an infrared small target detection method based on the fusion of dynamic prior and bidirectional interactive methods.

[0010] The technical solution of this invention is: an infrared small target detection method based on dynamic prior and bidirectional interactive fusion, characterized in that: the infrared small target detection method includes the following steps performed in sequence:

[0011] Step 1) Acquire the input infrared image and perform normalization preprocessing on the input infrared image;

[0012] Step 2) Construct a dynamic parameter prediction network to extract global statistical features from the input infrared image and predict dynamic prior parameters;

[0013] Step 3) Construct a dynamic multi-scale local contrast prior generation module using the dynamic prior parameters obtained in Step 2) and calculate the local contrast response, background suppression response and regional complexity response under multiple scale windows respectively;

[0014] Step 4) Using the responses at each scale obtained in Step 3), perform weighted fusion through a dynamic scale fusion mechanism to generate an initial saliency prior map;

[0015] Step 5) Construct an encoder-decoder network to extract multi-scale features from the input infrared image and obtain the decoded features. ;

[0016] Step 6) Construct a priori guided feature enhancement module using the initial saliency prior map obtained in Step 4) to enhance the spatial dimension and channel dimension of the encoder output features in Step 5) to obtain enhanced encoded features;

[0017] Step 7) Using the decoding features obtained in Step 5) and the enhanced coding features obtained in Step 6), construct a feature feedback correction module to dynamically correct the initial saliency prior map obtained in Step 4) and generate a refined saliency prior map;

[0018] Step 8) Construct a gated fusion module using the refined saliency prior map obtained in Step 7) to achieve adaptive fusion of encoded and decoded features, and obtain fused features;

[0019] Step 9) Generate a target probability map based on the fusion features obtained in Step 8) and output the infrared small target detection results;

[0020] Step 10) Construct a structural prior distillation mechanism to jointly supervise and optimize the initial saliency prior map, the refined saliency prior map, and the target probability map to update the network parameters.

[0021] Furthermore, the dynamic parameter prediction network described in step 2) includes: a convolutional feature extraction layer, a global average pooling layer, and a fully connected mapping layer, as detailed below:

[0022] First, a convolutional feature extraction layer is used to obtain the global features of the input image. Let the input infrared image be... First, global features are extracted using a lightweight convolutional network. :

[0023] ;

[0024] in, This represents a feature extraction function composed of multiple convolutional layers and global average pooling.

[0025] Based on this, a fully connected mapping layer is used. Predicting dynamic prior parameters:

[0026] ;

[0027] in, This represents the contrast enhancement coefficient, used to control the degree of amplification of the target response; This represents the background suppression coefficient, used to adjust the suppression strength of background interference; This represents the normalization adjustment coefficient, used to balance the response ratio between the target and the background. This represents the lower limit threshold for background fluctuations, used to enhance numerical stability.

[0028] Furthermore, in step 3), the dynamic multi-scale local contrast prior generation module uses multiple windows of different scales to perform saliency analysis on the target. The specific method is as follows:

[0029] The scale windows include: 3×3 windows, 5×5 windows, 7×7 windows, and 9×9 windows, which respectively obtain the local contrast response at the corresponding scale. The calculation formula is as follows:

[0030] ;

[0031] in, For dynamic contrast enhancement coefficient, To prevent small constants from being numerically unstable, The expression is defined as the contrast response. Indicates the maximum response in the domain context. This represents the result of Gaussian smoothing after scaling the input image.

[0032] Furthermore, in step 3), a background suppression response is constructed to distinguish between the real target and the highlight interference in the complex background. The specific steps are as follows:

[0033] ;

[0034] This indicates the definition of local differences at the current scale. Indicates a local maximum response. This represents the local average intensity over a window of scale s; meanwhile, for the neighborhood offset location, the following is defined:

[0035] ;

[0036] top right mark This indicates the selected neighborhood direction number, where the k value includes the 8 directions surrounding the center pixel; and it takes the maximum difference in the neighborhood. :

[0037] ;

[0038] Based on this, a dynamic background suppression coefficient is introduced. Constructing a suppression term :

[0039] ;

[0040] Define the degree of background fluctuation :

[0041] ;

[0042] This indicates that standard deviation is calculated and constrained by a dynamic threshold:

[0043] ;

[0044] Indicates the dynamic threshold range. Indicates the background fluctuation threshold. Indicates taking The larger of the two;

[0045] Ultimately, background suppression prior. for:

[0046] ;

[0047] in, This is the normalized adjustment coefficient.

[0048] Furthermore, the dynamic scale fusion mechanism in step 4) uses the Softmax function to obtain the weights for each scale, which are used to measure the local energy distribution of the target at the current scale. The specific steps are as follows:

[0049] First, calculate the deviation between the image and the local mean. :

[0050] ;

[0051] This represents the result of Gaussian smoothing after scaling the input image. This represents the local average intensity within a window of scale s; then, averaging is performed within this scale window.

[0052] ;

[0053] This indicates the degree of energy concentration at the current scale. This indicates that average pooling calculation is being performed. Represents a scale set;

[0054] At every scale Three types of responses were obtained respectively. They are stacked along the channel dimension and weighted using dynamic weights:

[0055] ;

[0056] in, These represent the contrast prior, background suppression prior, and scale response prior, respectively. The scale weights are predicted by dynamic parameters; then, the three types of responses are further refined:

[0057] ;

[0058] , , They represent based on Detailed results of the three types of responses For the Sigmoid function; , , , It is a convolutional network;

[0059] Finally, the three types of priors are fused to obtain a unified prior graph. :

[0060] ;

[0061] in, This indicates channel splicing.

[0062] Furthermore, the prior-guided feature enhancement module in step 6) includes a spatial attention enhancement unit and a channel attention enhancement unit. The core objective is to integrate the target saliency information contained in the prior image into the encoded features at the current scale, enabling the network to prioritize the locations and channels more likely to contain the target before feature fusion. The specific steps are as follows:

[0063] Let the input prior map be... First, it is adjusted to match the current encoded features using bilinear interpolation. The same spatial resolution is denoted as:

[0064] ;

[0065] in, Indicates an upsampling operation;

[0066] First, in the spatial dimension, through a convolution Using the sigmoid activation function, a spatial attention map is generated from the prior map:

[0067] ;

[0068] in This is used to depict the importance of each spatial location;

[0069] Subsequently, in the channel dimension, through a convolution The single-channel prior is mapped to a channel modulation tensor with the same number of encoded feature channels, and batch normalization is combined. Channel attention map obtained with Sigmoid activation:

[0070] ;

[0071] in, This tensor can assign different weights to different channels at each spatial location;

[0072] After obtaining spatial attention and channel attention, residual modulation is used to enhance the encoded features, resulting in the guided feature representation:

[0073] ;

[0074] in, This represents element-wise multiplication. Here, we use... In form rather than directly using As a scaling factor, it is mainly used to avoid excessive suppression of certain locations or channels, which would lead to the loss of effective detail information. Through this residual enhancement method, the prior map selectively amplifies the response of the potential target region without destroying the original feature structure.

[0075] Furthermore, the feature feedback correction module in step 7) constructs a joint feature representation using encoded and decoded features to perform reverse correction on the prior. The specific steps are as follows:

[0076] First, decode the features Compared with prior enhanced coding features By concatenating the features along the channel dimension, we obtain the joint feature representation:

[0077] ;

[0078] in, It integrates two types of information: one is high-level semantic information from the decoding branch, and the other is detailed information from the encoding branch that has been enhanced a priori.

[0079] Simultaneously, the input prior image is upsampled to the current scale, denoted as... Then, the joint features are combined with priors. Figure 1 Using the same input prior correction network, predict a prior compensation term:

[0080] ;

[0081] in, Represents the prior calibration subnetwork;

[0082] Finally, the updated prior graph is represented as follows:

[0083] ;

[0084] in, This indicates that the value will be cropped to... Within the range.

[0085] Furthermore, the gated fusion module in step 8) generates gate weights based on the refined saliency prior map, so that the fusion result retains both shallow structural details and deep semantic discrimination capabilities. The specific steps are as follows:

[0086] First, to facilitate subsequent fusion, the encoded and decoded features need to be mapped to a unified channel dimension; this is achieved by using... Projecting it using convolution, batch normalization, and ReLU activation:

[0087] ;

[0088] in, and Both indicate channel alignment operations. This represents the encoded features after mapping. This represents the decoded features after mapping;

[0089] Subsequently, the projected encoded features, decoded features, and updated prior image are concatenated along the channel dimension to construct the gated prediction input. :

[0090] ;

[0091] Then, a gated weight graph is generated through the gated subnetwork. :

[0092] ;

[0093] in, Represents the ReLU activation function. This represents the Sigmoid function. For the gating weight graph, It is not a single scalar, but a fine-grained weight tensor with spatial and channel dimensions, so it can adaptively determine which type of feature should be retained at each position and each channel;

[0094] After obtaining the gating weight map, the fusion form of the encoded features and decoded features is defined as follows: :

[0095] ;

[0096] Introducing a lightweight feature enhancement module The fusion result is then reorganized and refined, and its form can be expressed as follows:

[0097] ;

[0098] in, This represents a post-processing map consisting of depthwise separable convolutions and residual blocks. Depthwise separable convolutions can extract local structural features with low computational cost, while residual blocks help improve feature representation and mitigate information loss.

[0099] Furthermore, the structural prior distillation mechanism in step 10) includes pixel consistency distillation loss, edge consistency distillation loss, and feature consistency distillation loss, enabling the prior image to gradually learn more stable spatial responses, edge structures, and semantic feature representations in the detection branch during training, thereby improving the reliability and structural integrity of the prior. The specific steps are as follows:

[0100] Let the final predicted probability map be:

[0101] ;

[0102] When using it as a distillation target, a stopping gradient operation is employed:

[0103] ;

[0104] This indicates that gradient calculation is being performed, assuming the initial prior map is... Refine the prior diagram as The prior branch features are The predicted branch features are Structural prior distillation constrains the prior through three levels: pixel consistency, edge consistency, and feature consistency. The overall loss is defined as:

[0105] ;

[0106] in, The main detection loss consists of a binary cross-entropy loss function and a soft-IoU loss function. It is a priori and multi-branch auxiliary supervision loss. This represents the overall loss from structural prior distillation, which defines the constraints on the priors. , They represent , The weighting coefficients.

[0107] Beneficial effects:

[0108] This invention pioneers a dynamic multi-scale local comparison prior generation mechanism, abandoning the traditional manual fixed parameter design. It adaptively generates a complete set of control parameters through global image features, which can accurately adapt to various complex low-altitude scenes such as smooth sky background, building clutter background, and sea surface heat source background. It completely solves the technical problems of poor adaptation of static prior and high false alarm rate in complex scenes, and greatly improves the model's scene generalization ability.

[0109] This invention constructs a two-way interactive fusion closed-loop system of prior and deep features, which is different from the existing one-way prior injection mode. It realizes two-way collaborative optimization of prior-guided feature enhancement and deep feature reverse correction of prior, effectively corrects the defects of initial prior noise pseudo response, structural bias and other defects, eliminates error accumulation in the training process, and significantly improves the accuracy and completeness of small target feature extraction.

[0110] This invention designs a multi-layered prior distillation constraint mechanism, constructing joint loss constraints from three dimensions: pixel grayscale, edge contour, and high-level features. This comprehensively optimizes the target detection structure, effectively solving the problems of fragmented small target contours, positioning offset, and target breakage in existing technologies, and significantly improving the detection rate of micro-UAVs in extremely low signal-to-noise ratio scenarios.

[0111] This invention adopts a lightweight network architecture design with only 1.837M model parameters and an inference frame rate of up to 132.53 FPS. Compared with existing mainstream algorithms, the number of parameters is greatly reduced and the real-time performance is significantly improved. It can be stably deployed on low-computing-power hardware such as infrared PTZ cameras, airborne embedded terminals, and low-altitude security monitoring equipment, making it highly practical for engineering applications.

[0112] This invention exhibits excellent detection performance. On three authoritative publicly available infrared small target datasets—NUAA-SIRST, NUDT-SIRST, and IRSTD-1K—its core metrics, such as intersection-over-union ratio, detection accuracy, and false alarm rate, are significantly superior to traditional algorithms and existing deep learning algorithms. It can effectively adapt to infrared small target detection scenarios with different scales, signal-to-noise ratios, and backgrounds, demonstrating outstanding robustness and generalization ability.

[0113] This invention features end-to-end intelligent detection without the need for manual parameter adjustment. It adapts to various complex low-altitude monitoring environments, enabling all-weather, fully automated infrared UAV target early warning and detection, and is suitable for large-scale low-altitude security monitoring applications. Attached Figure Description

[0114] Figure 1 The overall DPIF-Net network architecture diagram provided by this invention.

[0115] Figure 2 The diagram shows the architecture of the dynamic prior generation module provided by this invention.

[0116] Figure 3 The multi-scale prior network structure diagram provided by this invention.

[0117] Figure 4 The priori-guided special detection enhancement network structure diagram provided by the present invention.

[0118] Figure 5 The feature feedback verification prior network structure diagram provided by this invention.

[0119] Figure 6 The diagram shows the structure of the gating fusion network provided by this invention.

[0120] Figure 7 This is a visualization of comparative experiments related to the method described in this invention. Detailed Implementation

[0121] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0122] like Figure 1 As shown, the infrared small target detection method based on dynamic prior and bidirectional interactive fusion provided by this invention includes the following steps performed in sequence:

[0123] 1) Acquire the input infrared image and perform normalization preprocessing on the image to map the pixel values ​​to a uniform numerical range, so as to reduce the impact of different imaging conditions on network training and improve the stability and convergence speed of model training.

[0124] 2) such as Figure 2As shown, a dynamic prior generation module is constructed. First, a convolutional feature extraction network is used to extract global features from the input infrared image. Global average pooling is used to obtain the overall statistical information of the image. Then, a fully connected mapping layer is used to predict dynamic prior parameters, including the target contrast enhancement coefficient, background suppression coefficient, normalization adjustment coefficient, and background fluctuation threshold parameter, thereby achieving adaptive prior modeling for different scenarios.

[0125] 3) such as Figure 3 As shown, a dynamic multi-scale local contrast prior generation module is constructed based on the dynamic prior parameters obtained in step 2. Local saliency analysis is performed on the input image using 3×3, 5×5, 7×7, and 9×9 multi-scale windows, calculating the local contrast response, background difference response, and regional complexity response at each scale to enhance the distinguishing ability between the target region and the background region.

[0126] 4) Using the local contrast response, background suppression response, and regional complexity response obtained at each scale, a weighted fusion is performed through a dynamic scale fusion mechanism, and combined with a convolutional thinning network to generate a unified saliency prior map. This prior map can highlight potential target regions and suppress complex background interference, providing prior guidance information for subsequent feature extraction.

[0127] 5) Construct an encoder-decoder network to perform multi-level feature extraction on the input infrared image. The encoder extracts local detail features and high-level semantic features of the target layer by layer, while the decoder gradually restores the spatial resolution and fuses feature information from different levels to obtain a feature representation with rich semantic expressive power.

[0128] 6) For example Figure 4 As shown, a priori-guided feature enhancement module is constructed. The saliency prior map generated in step 4 is adjusted to the same resolution as the encoded features, and attention weights are generated from both the spatial and channel dimensions to adaptively enhance the encoder output features, thereby improving the network's attention to small infrared target regions.

[0129] 7) For example Figure 5 As shown, a feature feedback correction module is constructed. The high-level semantic features obtained during the decoding stage are fused with the enhanced prior encoded features, and a prior compensation term is predicted based on the saliency prior map at the current scale. This dynamically corrects the initial saliency prior map, resulting in a more accurate refined saliency prior map.

[0130] 8) For example Figure 6As shown, a gating fusion module is constructed. First, the encoded and decoded features are channel aligned. Then, the aligned encoded and decoded features, along with the refined saliency prior map, are input into the gating network to generate fine-grained gating weights. Based on these gating weights, the encoded and decoded features are adaptively fused, and the fusion result is further optimized using a lightweight feature enhancement module to obtain the final detection features.

[0131] 9) Construct a structural prior distillation mechanism to jointly supervise and optimize the initial saliency prior map, the refined saliency prior map, and the target prediction results. Through pixel consistency constraints, edge consistency constraints, and feature consistency constraints, guide the prior branch and the detection branch to learn collaboratively, improve the model's detection accuracy and robustness for small infrared targets in complex backgrounds, and finally output the target probability map and obtain the detection results.

[0132] An application example of the infrared small target detection method based on dynamic prior and bidirectional interactive fusion provided by the present invention is given below:

[0133] This embodiment relates to the application of an infrared small target detection method based on dynamic prior and bidirectional interactive fusion. This method is applicable to scenarios such as low-altitude security, border patrol, maritime surveillance, nighttime target search, and infrared UAV detection in complex backgrounds, and is particularly suitable for infrared image target recognition tasks with extremely small target size, strong background interference, and low signal-to-noise ratio.

[0134] (1) In this embodiment, three public datasets, NUAA-SIRST, NUDT-SIRST and IRSTD-1k, are selected as application verification objects. The targets in the NUAA-SIRST dataset are mostly very weak point targets, which is suitable for verifying the basic detection capability of the method under low signal-to-noise ratio conditions; the NUDT-SIRST dataset contains more diverse target sizes and complex backgrounds, which is suitable for verifying the multi-scale adaptability of the method; the IRSTD-1k dataset has more complex background types, including clouds, buildings, trees, reflective water surfaces and other interference scenes, which is suitable for verifying the generalization ability and robustness of the method in complex environments. The basic information of the three public datasets is shown in Table 1.

[0135] Table 1. Basic information about the three types of public datasets.

[0136]

[0137] (2) First, the infrared image to be detected is input into the dynamic multi-scale local contrast prior generation module. This module automatically constructs a corresponding saliency prior map based on the local grayscale changes and neighborhood background differences of the input image, which is used to highlight potential target areas and suppress large-area background responses. Unlike traditional static priors, this prior map has dynamic generation characteristics and can adaptively adjust the response distribution according to different scenes. Therefore, it has good adaptability to UAV targets in the sky, sea, fields, forests, and building backgrounds.

[0138] Subsequently, the generated prior map is bidirectionally fused with the deep features extracted by the encoder. Specifically, on the one hand, the prior map guides feature enhancement, enabling the network to prioritize high-response regions in the shallow stages and avoid background clutter interference; on the other hand, the semantic features extracted in the decoding stage act inversely on the prior map, correcting spurious responses, broken responses, or excessively strong background responses in the prior. Through this bidirectional interaction, the prior is no longer a fixed, static guiding signal, but can be continuously improved as semantic features are gradually refined, thereby enhancing the accuracy and completeness of target region localization.

[0139] (3) In the feature fusion stage, this embodiment further introduces a gating fusion mechanism. This mechanism generates a gating weight map based on the updated prior map, encoded features, and decoded features, and accordingly performs position-by-position and channel-by-channel adaptive weighting of shallow detail features and deep semantic features. The advantage of this approach is that when the target region has weak details but clear positional information, the model can rely more on the spatial details preserved in the encoding stage; when the background is complex and local texture interference is strong, the model can make more use of the semantic information in the decoding stage for suppression. This mechanism makes feature fusion no longer a simple concatenation or fixed addition, but dynamically selects more favorable feature representations based on prior indications and contextual information.

[0140] (4) During the training phase, this embodiment also introduces a structural prior distillation method to constrain the consistency between the prior branch and the detection branch. This constraint not only focuses on the response alignment at the pixel level, but also further considers the consistency of edge structures and the consistency of high-level feature semantics. In this way, the prior image can gradually learn a more stable and complete target structure representation during training, reducing false detections and false negatives caused by background brightness or local noise. This strategy is particularly important for infrared weak targets, because the target itself has a sparse structure, and without structural constraints, the prior image is prone to local breaks or blurred boundaries.

[0141] (5) During testing, the trained model is deployed at the infrared image input end to perform frame-by-frame detection on the real-time acquired images. The model outputs the spatial location of the target, the response probability, and the corresponding segmentation results, which can be used for subsequent alarm triggering, target tracking, or airspace monitoring decisions.

[0142] Table 2. Comparison of the proposed model with the reference method on the NUAA-SIRST, NUDT-SIRST, and IRSTD-1K datasets in terms of IoU (%), nIoU (%), Pd (%), and Fa (%). Comparison results table on )

[0143]

[0144] (6) Experimental results show that the proposed method exhibits superior detection performance on three public datasets, significantly enhancing the response to small targets in complex backgrounds while effectively suppressing background clutter and false target interference. The comparison results of the proposed method and the reference method on different datasets are shown in Table 2, where ↑ indicates that the larger the result, the better, and ↓ indicates that the smaller the result, the better. Red represents the best result in this column, and blue represents the second best result in this column. After introducing prior guidance, the proposed method can more accurately focus on weak target regions. Therefore, Pd increased from 84.72% to 96.61% on NUAA-SIRST, and Fa decreased from 66.73 to 7.95 on NUDT-SIRST, demonstrating better target detection and false alarm suppression capabilities.

[0145] (7) This paper further conducted visualization experiments on the IRSTD-1K dataset. The experimental results are as follows: Figure 7 As shown in the diagram, red circles represent correctly detected targets, yellow circles represent false positives, green circles represent missed targets, and the blue box represents a magnified view of the red circles. GT represents the ground truth label. It can be seen that WSLCM's detection results have significant missed and false positives, and the binarized target morphology differs greatly from the ground truth label, often appearing as isolated bright spots and failing to preserve target structural information. ALCNet and U-Net show improvements over WSLCM, but still have a small number of false positives or missed targets. In contrast, the proposed method can more accurately locate the target region, and the segmentation results are closer to the ground truth label, indicating that dynamic prior modeling and bidirectional interactive fusion can effectively improve target localization accuracy and structural integrity.

Claims

1. An infrared small target detection method based on dynamic prior and bidirectional interactive fusion, characterized in that: The infrared small target detection method includes the following steps performed in sequence: Step 1) Acquire the input infrared image and perform normalization preprocessing on the input infrared image; Step 2) Construct a dynamic parameter prediction network to extract global statistical features from the input infrared image and predict dynamic prior parameters; Step 3) Construct a dynamic multi-scale local contrast prior generation module using the dynamic prior parameters obtained in Step 2) and calculate the local contrast response, background suppression response and regional complexity response under multiple scale windows respectively; Step 4) Using the responses at each scale obtained in Step 3), perform weighted fusion through a dynamic scale fusion mechanism to generate an initial saliency prior map; Step 5) Construct an encoder-decoder network to extract multi-scale features from the input infrared image and obtain the decoded features. ; Step 6) Construct a priori guided feature enhancement module using the initial saliency prior map obtained in Step 4) to enhance the spatial dimension and channel dimension of the encoder output features in Step 5) to obtain enhanced encoded features; Step 7) Using the decoding features obtained in Step 5) and the enhanced coding features obtained in Step 6), construct a feature feedback correction module to dynamically correct the initial saliency prior map obtained in Step 4) and generate a refined saliency prior map; Step 8) Construct a gated fusion module using the refined saliency prior map obtained in Step 7) to achieve adaptive fusion of encoded and decoded features, and obtain fused features; Step 9) Generate a target probability map based on the fusion features obtained in Step 8) and output the infrared small target detection results; Step 10) Construct a structural prior distillation mechanism to jointly supervise and optimize the initial saliency prior map, the refined saliency prior map, and the target probability map to update the network parameters.

2. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 1, characterized in that: The dynamic parameter prediction network described in step 2) includes: a convolutional feature extraction layer, a global average pooling layer, and a fully connected mapping layer, as detailed below: First, a convolutional feature extraction layer is used to obtain the global features of the input image. Let the input infrared image be... First, global features are extracted using a lightweight convolutional network. : ； in, This represents a feature extraction function composed of multiple convolutional layers and global average pooling. Based on this, a fully connected mapping layer is used. Predicting dynamic prior parameters: ； in, This represents the contrast enhancement coefficient, used to control the degree of amplification of the target response; This represents the background suppression coefficient, used to adjust the suppression strength of background interference; This represents the normalization adjustment coefficient, used to balance the response ratio between the target and the background. This represents the lower limit threshold for background fluctuations, used to enhance numerical stability.

3. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 1, characterized in that: Step 3) uses the dynamic multi-scale local contrast prior generation module to perform saliency analysis on the target using multiple windows of different scales. The specific method is as follows: The scale windows include: 3×3 windows, 5×5 windows, 7×7 windows, and 9×9 windows, which respectively obtain the local contrast response at the corresponding scale. The calculation formula is as follows: ； in, For dynamic contrast enhancement coefficient, To prevent small constants from being numerically unstable, The expression is defined as the contrast response. Indicates the maximum response in the domain context. This represents the result of Gaussian smoothing after scaling the input image.

4. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 1, characterized in that: Step 3) further constructs a background suppression response to distinguish between the real target and the highlight interference in the complex background. The specific steps are as follows: ； This indicates the definition of local differences at the current scale. Indicates a local maximum response. This represents the local average intensity over a window of scale s; meanwhile, for the neighborhood offset location, the following is defined: ； top right mark This indicates the selected neighborhood direction number, where the k value includes the 8 directions surrounding the center pixel; and it takes the maximum difference in the neighborhood. : ； Based on this, a dynamic background suppression coefficient is introduced. Constructing a suppression term : ； Define the degree of background fluctuation : ； This indicates that standard deviation is calculated and constrained by a dynamic threshold: ； Indicates the dynamic threshold range. Indicates the background fluctuation threshold. Indicates taking The larger of the two; Ultimately, background suppression prior. for: ； in, This is the normalized adjustment coefficient.

5. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 1, characterized in that: The dynamic scale fusion mechanism in step 4) uses the Softmax function to obtain the weights for each scale, which are used to measure the local energy distribution of the target at the current scale. The specific steps are as follows: First, calculate the deviation between the image and the local mean. : ； This represents the result of Gaussian smoothing after scaling the input image. This represents the local average intensity within a window of scale s; then, averaging is performed within this scale window. ； This indicates the degree of energy concentration at the current scale. This indicates that average pooling calculation is being performed. Represents a scale set; At every scale Three types of responses were obtained respectively. They are stacked along the channel dimension and weighted using dynamic weights: ； in, These represent the contrast prior, background suppression prior, and scale response prior, respectively. The scale weights are represented by those predicted from the dynamic parameters; then, the three types of responses are further refined: ； , , They represent based on Detailed results of the three types of responses For the Sigmoid function; , , , It is a convolutional network; Finally, the three types of priors are fused to obtain a unified prior graph. : ； in, This indicates channel splicing.

6. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 1, characterized in that: Step 6) includes a priori-guided feature enhancement module comprising a spatial attention enhancement unit and a channel attention enhancement unit. Its core objective is to integrate target saliency information from the prior image into the encoded features at the current scale, enabling the network to prioritize locations and channels more likely to contain the target before feature fusion. The specific steps are as follows: Let the input prior map be... First, it is adjusted to match the current encoded features using bilinear interpolation. The same spatial resolution is denoted as: ； in, Indicates an upsampling operation; First, in the spatial dimension, through a convolution Using the sigmoid activation function, a spatial attention map is generated from the prior map: ； in This is used to depict the importance of each spatial location; Subsequently, in the channel dimension, through a convolution The single-channel prior is mapped to a channel modulation tensor with the same number of encoded feature channels, and batch normalization is combined. Channel attention map obtained with Sigmoid activation: ； in, This tensor can assign different weights to different channels at each spatial location; After obtaining spatial attention and channel attention, residual modulation is used to enhance the encoded features, resulting in the guided feature representation: ； in, This indicates element-wise multiplication.

7. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 1, characterized in that: The feature feedback correction module in step 7) constructs a joint feature representation using encoded and decoded features to perform reverse correction on the prior. The specific steps are as follows: First, decode the features Compared with prior enhanced coding features By concatenating the features along the channel dimension, we obtain the joint feature representation: ； in, It integrates two types of information: one is high-level semantic information from the decoding branch, and the other is detailed information from the encoding branch that has been enhanced a priori. Simultaneously, the input prior image is upsampled to the current scale, denoted as... Then, the joint features and the prior map are input into the prior correction network to predict a prior compensation term: ； in, Represents the prior calibration subnetwork; Finally, the updated prior graph is represented as follows: ； in, This indicates that the value will be cropped to... Within the range.

8. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 7, characterized in that: The gated fusion module in step 8) generates gate weights based on the refined saliency prior map, so that the fusion result retains both shallow structural details and deep semantic discrimination capabilities. The specific steps are as follows: First, to facilitate subsequent fusion, the encoded and decoded features need to be mapped to a unified channel dimension; this is achieved by using... Projecting it using convolution, batch normalization, and ReLU activation: ； in, and Both indicate channel alignment operations. This represents the encoded features after mapping. This represents the decoded features after mapping; Subsequently, the projected encoded features, decoded features, and updated prior image are concatenated along the channel dimension to construct the gated prediction input. : ； Then, a gated weight graph is generated through the gated subnetwork. : ； in, Represents the ReLU activation function. This represents the Sigmoid function. For the gating weight graph, It is not a single scalar, but a fine-grained weight tensor with spatial and channel dimensions, so it can adaptively determine which type of feature should be retained at each position and each channel; After obtaining the gating weight map, the fusion form of the encoded features and decoded features is defined as follows: : ； Introducing a lightweight feature enhancement module The fusion result is then reorganized and refined, and its form can be expressed as follows: ； in, This represents a post-processing map consisting of depthwise separable convolutions and residual blocks.

9. The infrared small target detection method based on dynamic prior and bidirectional interactive fusion according to claim 8, characterized in that: The structural prior distillation mechanism in step 10) includes pixel consistency distillation loss, edge consistency distillation loss, and feature consistency distillation loss. This allows the prior image to gradually learn more stable spatial responses, edge structures, and semantic feature representations in the detection branch during training, thereby improving the reliability and structural integrity of the prior. The specific steps are as follows: Let the final predicted probability map be: ； When using it as a distillation target, a stopping gradient operation is employed: ； This indicates that gradient calculation is being performed, assuming the initial prior map is... Refine the prior diagram as The prior branch features are The predicted branch features are Structural prior distillation constrains the prior through three levels: pixel consistency, edge consistency, and feature consistency. The overall loss is defined as: ； in, The main detection loss consists of a binary cross-entropy loss function and a soft-IoU loss function. It is a priori and multi-branch auxiliary supervision loss. This represents the overall loss from structural prior distillation, which defines the constraints on the priors. , They represent , The weighting coefficients.

Images