An infrared small target detection system and method fusing physical saliency features

Abstract

The application discloses an infrared small target detection system and method fusing physical saliency features, relates to the technical field of infrared small target detection, and solves the problems that the existing infrared small target detection method causes high false alarm and high missing detection in a complex dynamic background and the existing deep learning model has high false alarm rate and insufficient robustness when detecting infrared small targets lacking texture and shape information. The method comprises the following steps: acquiring a single frame of an infrared image to be detected, inputting the infrared image into a pre-trained network model, outputting a prediction bounding box, a class confidence and physical saliency features of a target in parallel through the network model, fusing the class confidence and the physical saliency features through a post-processing fusion module, obtaining a comprehensive confidence, filtering the comprehensive confidence, and realizing infrared small target detection. The application eliminates the dependence on continuous frame caching and processing, realizes zero initialization delay and frame-level real-time detection, and meets the application scenarios with extremely high real-time requirements such as early warning and guidance.

Description

Technical Field

[0001] This invention relates to the field of infrared small target detection technology, specifically to an infrared small target detection system and method that integrates physical saliency features. Background Technology

[0002] In the field of infrared small target detection, traditional detection algorithms based on single-frame image features (such as contrast and gradient) can describe physical saliency, but lack high-level semantic understanding. When the target shape is varied or similar to the background, the detection rate and robustness are insufficient. General target detection models, such as YOLOv8, were not designed for pixel-level infrared small targets. They are better at using semantic information such as the target's shape, parts, and context to determine the target type. However, infrared small targets lack this information. Therefore, YOLOv8 has a high false alarm rate and low recall rate when detecting infrared small targets, and it is also prone to overfitting during training.

[0003] Infrared imaging systems operate by sensing the difference in thermal radiation between a target and its background, enabling them to conduct observations in all weather conditions, including nighttime, fog, haze, and dust storms. Therefore, infrared small target detection is widely used in national defense (e.g., early warning and guidance), public safety (e.g., border surveillance and search and rescue), and civilian applications (e.g., power line inspection and forest fire prevention).

[0004] The core challenge of infrared imaging technology stems from the characteristics of the target itself:

[0005] 1. The target occupies only a few pixels in the image (e.g., 3×3 to 10×10 pixels) and lacks detailed shape and texture information that can be identified.

[0006] 2. The target signal is weak, and the signal-to-noise ratio (SNR) and contrast ratio (SCR) are usually very low, making it easy to be submerged in complex background noise and clutter.

[0007] 3. Unlike targets in natural images, infrared small targets lack semantic features such as rich colors, clear outlines, or fixed structures.

[0008] These characteristics lead to poor performance of traditional general object detection methods based on appearance semantics, and high false alarm rates have become a long-standing technical bottleneck.

[0009] Existing technology 1, Low-Rank Tensor Quantum Space Learning for Infrared Small Target Detection, published in Beijing in 2024, DOI: 10.16337 / j.1004-9037.2025.02.006, proposes a low-rank tensor quantum space learning algorithm (TMSL). Addressing the low robustness of current spatiotemporal algorithms to various noise clutter interferences and complex scene combinations, it utilizes the powerful adaptive representation capability of subspaces to represent complex high-dimensional image components as low-rank tensor features with unique and important information. It achieves relatively good detection performance in infrared videos with complex backgrounds. However, low-rank decomposition works well with simple backgrounds. But when faced with complex dynamic backgrounds, background changes no longer satisfy a simple low-rank model, leading to the decomposed "sparse components" containing a large amount of background residuals, resulting in extremely high false alarm rates. Methods typically assume that the target is "sparse" and the noise is independent and identically distributed. However, in reality, background clutter can easily be misidentified as a target; and weak target signals may be submerged by noise, failing to meet strong sparsity, thus leading to missed detections. Its performance is highly dependent on the precise tuning of the regularization parameters, resulting in insufficient robustness.

[0010] Existing technology 2, a method for detecting weak infrared targets in complex backgrounds based on spatiotemporal three-dimensional convolutional networks, was published in Changchun in 2025, with DOI 1007-2780(2025)11-1688-12. This method proposes a method for detecting weak targets. It proposes a feature extraction backbone network that combines two-dimensional convolution and three-dimensional convolution, and combines spatial texture features and inter-frame motion features to achieve collaborative perception of target structure and temporal changes. It introduces an asymmetric attention mechanism for feature fusion to increase the preservation of texture and position information. Finally, the detection result is calculated through a point regression loss function.

[0011] Existing technology 3, a lightweight infrared small target detection based on point-to-point regression, published in Xi'an, 2026 (DOI: 10.3788 / gzxb20265502.0210002), proposes an asymmetric full-dimensional dynamic feature extraction module, constructing structures such as channel shuffling and deep convolution to efficiently enhance key target information while suppressing background interference. Then, a high-resolution cross-feature enhancement module is designed, utilizing the cross-fusion of multiple high-resolution feature extraction branches to simplify the traditional multi-level convolution stacking structure, effectively mining high-level semantic features while preserving target detail information. However, in this type of deep learning-based infrared point target detection, the model's predictions often fail when the target undergoes drastic maneuvers (sharp turns, speed changes), is briefly occluded, or suddenly appears / disappears, leading to trajectory interruptions or false alarms.

[0012] Therefore, this invention aims to provide an infrared small target detection solution based on a single-frame image, which deeply integrates physical saliency priors and deep learning semantic features. Specifically, the prediction of physical saliency features (GNSD) is embedded into a deep neural network as a learnable task. The deeper purpose is to solve the problem of insufficient detection performance of traditional detection methods or other fusion methods and deep methods. Summary of the Invention

[0013] This invention addresses the problems of high false alarm rates and high false negative rates in traditional infrared small target detection methods under complex dynamic backgrounds due to invalid model assumptions, as well as the high false alarm rate and insufficient robustness of existing deep learning models (such as YOLOv8) when detecting infrared small targets lacking texture and shape information. It provides an infrared small target detection system and method that integrates physical saliency features.

[0014] An infrared small target detection system that integrates physical saliency features, the detection system includes a GNSD_YOLO network model, which consists of a feature extraction backbone network, a feature pyramid network, a target detection head, and a post-processing fusion module;

[0015] The target detection head consists of three parallel bounding box regression heads, a target classification head, and a GNSD feature prediction head;

[0016] The feature extraction backbone network extracts multi-scale features from the input infrared image and performs bidirectional cross-scale fusion through the feature pyramid network to output enhanced features.

[0017] The enhanced features are decoded in parallel through the bounding box regression head, the target classification head, and the GNSD feature prediction head, respectively, to output the bounding box coordinates, class confidence, and physical saliency features of the target.

[0018] The post-processing fusion module fuses the category confidence and physical saliency features to obtain a comprehensive confidence score, and filters the comprehensive confidence score to achieve infrared small target detection.

[0019] This invention also provides an infrared small target detection method that integrates physically saliency features. This method is implemented using the aforementioned infrared small target detection system that integrates physically saliency features, and specifically comprises the following steps:

[0020] Step S1. Acquire a single-frame infrared image to be detected;

[0021] Step S2. Input the infrared image into the pre-trained GNSD_YOLO network model; through the GNSD_YOLO network model, output the predicted bounding box, class confidence and physical saliency features of the target in parallel;

[0022] Step S3. The semantic category confidence and physical saliency features are fused through the post-processing fusion module to obtain a comprehensive confidence. The comprehensive confidence is then filtered to achieve infrared small target detection.

[0023] The beneficial effects of this invention are:

[0024] The method of this invention effectively integrates salient features reflecting the essential physical properties of the target (such as local contrast, gradient difference, and regional compactness) with the semantic perception ability of the deep learning model, thereby improving the model's ability to distinguish between real targets and background clutter and significantly reducing the false alarm rate in complex scenarios.

[0025] This invention modifies the network structure to enable the model to directly predict a composite feature value that reflects both the physical salience of the target and the clutter of the background (e.g., ...). When making a decision, this predicted value is fused with the model's original classification confidence, thereby simultaneously considering semantics. and physical salience characteristics To make the final judgment, filter out those false alarm boxes that are semantically ambiguous but physically insignificant.

[0026] Compared to traditional methods such as low-rank decomposition, this approach avoids the high false alarm rate and high missed detection rate problems caused by invalid model assumptions (non-low-rank background, non-sparse target). By introducing physically saliency features as additional supervision and decision-making criteria, it effectively suppresses the model from misclassifying background clutter as targets, thereby significantly reducing the false alarm rate while maintaining high recall.

[0027] This invention is based entirely on inference from a single frame of image. This eliminates the reliance on continuous frame buffering and processing, achieving zero initialization latency and frame-level real-time detection, meeting the requirements of applications with extremely high real-time demands, such as early warning and guidance. Furthermore, this invention is unaffected by violent target maneuvers, brief obstructions, or platform movement.

[0028] This invention adds only a very lightweight prediction branch (GNSD feature prediction head) to the original YOLOv8 detection head, and the additional number of parameters and computational overhead introduced are negligible. Compared with solutions that redesign complex backbone networks or add a large number of modules, this invention maintains the advantages of the original model in terms of efficiency and lightweight, is easy to deploy on embedded devices or edge computing platforms, and has high engineering practical value. Attached Figure Description

[0029] Figure 1This is a structural block diagram of an infrared small target detection system that integrates physical saliency features according to the present invention;

[0030] Figure 2 This is a block diagram illustrating the principle of the improved deep learning detection model GNSD_YOLO in this invention.

[0031] Figure 3 This is a flowchart of an infrared small target detection method that integrates physical saliency features, as described in this invention.

[0032] Figure 4 A flowchart for calculating physical saliency features;

[0033] Figure 5 The image is a schematic diagram of an infrared image from the commonly used infrared dataset IRSTD-1K; where (a) is a small target next to a cluster of urban buildings; (b) is a small target with sparse interference in the night sky; (c) is a small target in complex mountainous terrain; (d) is a small target camouflaged by cloud cover; (e) is a small target hidden by vegetation; and (f) is a small target in flat terrain.

[0034] Figure 6 A schematic diagram illustrating the decrease in the loss function of the GNSD detection channel;

[0035] Figure 7 A schematic diagram illustrating the decrease in the loss function of the bounding box detection channel;

[0036] Figure 8 This is a comparison of the detection results of GNSD_YOLO and YOLOv8. The left column is the original image, the middle column is the prediction result of GNSD_YOLO, and the right column is the prediction result of YOLOv8. The data marked in the image is the confidence score of the target detection. Detailed Implementation

[0037] Specific Implementation Method 1: Combination Figure 1 and Figure 2 This embodiment describes an infrared small target detection system that integrates physical saliency features. The detection system includes an image input interface and an improved deep learning network model (GNSD_YOLO network model). The GNSD_YOLO network model consists of a feature extraction backbone network, a neck network (feature pyramid network), a target detection head, and a post-processing fusion module.

[0038] Infrared images are input into a feature extraction backbone network via an image input interface. This backbone network extracts multi-scale features, reusing the structure of YOLOv8. Then, a feature pyramid network performs bidirectional cross-scale fusion, concatenating deep semantic information with shallow detail features to enhance the detection capabilities and form more robust detection features. Figure 2As shown, the feature pyramid network constructs a bidirectional feature pyramid through cascaded cross-stage local network (C2f) modules, bottleneck convolution (BC) modules, and upsampling (US) modules. The C2f module captures multi-scale contextual semantic information, and the BC module compresses channels to reduce computation. The US operation restores spatial resolution, ultimately outputting feature tensors at different levels. Finally, the enhanced features are fed into the target detection head, which consists of three parallel bounding box regression heads, a target classification head, and a GNSD feature prediction head. The enhanced features are decoded through multi-layer convolution (Conv, Conv2d) operations to obtain the target's predicted bounding box coordinates (bbox), class confidence (conf), and physical saliency features (gnsd).

[0039] In this embodiment, a GNSD feature prediction head is added inside the target detection head, in addition to the original bounding box regression head and target classification head. After modification, the model architecture is as follows: the traditional YOLO only includes the predicted bounding box position. and category confidence (conf) output channel, The coordinates of the center point of the target. , The width and height are specified. In this embodiment, a GNSD output channel (GNSD feature prediction head) is added in parallel to output the physical saliency features predicted by the model; the overall target detection head is reconstructed into a three-task parallel output architecture.

[0040] In this embodiment, the GNSD feature prediction head of the GNSD_YOLO network model shares the same set of multi-scale feature maps from the feature pyramid network as input as the original branch. It is usually composed of several lightweight convolutional layers, and finally reduces the number of channels to 1 through a linear layer (or 1×1 convolution), and uses a sigmoid activation function at the output to constrain the predicted value to a reasonable range (such as [0, 1]), making it an interpretable saliency score.

[0041] In this embodiment, during the training phase of the GNSD_YOLO network model, the overall loss is determined by the detection loss function. and GNSD feature regression loss function The detection loss function is obtained through joint weighting. It is composed of the bounding box regression loss, intersection-over-union (IoU) loss, and class classification loss, weighted according to certain criteria. In this implementation, a separate GNSD feature regression loss function is used for the GNSD output channel, calculated using a traditional regression function (Smooth L1 Loss), expressed as follows:

[0042]

[0043] In the formula, It is usually set to 1. The coordinates of the center point of the target.

[0044] Therefore, the total loss function of the GNSD_YOLO network model as follows:

[0045]

[0046] in, It is a hyperparameter that balances task weights. The ground truth label of the GNSD feature regression loss function is calculated by applying an algorithm to each ground truth bounding box on the offline training dataset.

[0047] In this embodiment, through end-to-end joint training, the GNSD_YOLO network model learns how to locate and classify targets, and also learns how to accurately predict the physical salience of a bounding box, providing the model with prior knowledge of the essential features of infrared small targets.

[0048] Specific Implementation Method Two: Combination Figure 3 and Figure 4 This embodiment describes a detection method for an infrared small target detection system that integrates physical saliency features, as described in Specific Embodiment 1. This method is implemented through the following steps:

[0049] Step S1: Acquire a single-frame infrared image to be detected;

[0050] Step S2: Input the infrared image into the pre-trained improved deep learning network model GNSD_YOLO; through the GNSD_YOLO network model, output the predicted bounding box coordinates, class confidence and physical saliency features of the target in parallel;

[0051] In this embodiment, the bounding box regression head outputs the predicted bounding box coordinates, namely: the target position and size. The target classification head outputs the confidence level of the category containing the target within the bounding box; the GNSD feature prediction head outputs a feature reflecting the physical salience of the region.

[0052] In this embodiment, the physical salience feature This method integrates low-level visual features such as gradient contrast, local contrast, and regional compactness of the target region, and incorporates a background clutter penalty mechanism to calculate a scalar value. Its physical meaning is clear, effectively distinguishing the real target from background clutter, making it suitable for prediction using the YOLO model, and ultimately adjusting the detection results. The final physical saliency feature... The calculation process is as follows Figure 4 As shown, it includes the following steps:

[0053] Step S21: Input and preprocessing; acquire a single frame infrared image and candidate bounding boxes ,in, The coordinates of the center point of the target. , The width and height are given. The infrared image is calculated. gradient magnitude plot and contrast enhancement image ;

[0054] Step S22: Region extraction; extracting regions from the infrared image (original image) respectively. Gradient magnitude diagram and contrast enhancement image Extracting from bounding boxes A defined set of pixels for the target region. Simultaneously, a ring-shaped region is generated, extending outwards by a certain proportion (e.g., 50%) from the target region as the background pixel set, used to calculate the multiple feature values ​​of the target in subsequent calculations.

[0055] Step S23: Parallel computation of multiple features; based on the pixels of the target region and the background region, calculate the following three types of features:

[0056] Gradient contrast features Calculate the difference between the target region and the background region on the gradient magnitude map. The gradient contrast feature... The mean gradient magnitude of the target region on the gradient magnitude map reflects the average gradient difference level of the target itself, and is calculated using the following formula:

[0057]

[0058] in, and These are the mean and standard deviation of the gradient in the target region, respectively. and These are the mean and standard deviation of the gradient in the background region, respectively. The area of ​​the target region. To prevent division by zero, use a very small constant.

[0059] Local contrast features : CLAHE-based enhanced graph The contrast difference between the target and background regions is calculated using the following formula:

[0060]

[0061] in, and These are the mean values ​​of the target and background on the contrast map, respectively.

[0062] Regional compactness characteristics : Calculate the target region in the original image The uniformity of grayscale is defined as the ratio of the mean to the standard deviation, and is calculated using the following formula:

[0063]

[0064] in, and The mean and standard deviation of the grayscale values ​​of the target region in the original image.

[0065] Step S24: Calculate the background clutter penalty factor; calculate the clutter level of the background region and generate a background clutter penalty factor. This factor value ranges from 0 to 1. Background clutter It is determined by a combination of gradient dispersion and contrast uniformity in the background area:

[0066]

[0067] Then, the Sigmoid function is used to... Mapped to a penalty factor: The more cluttered the background, The higher the value, the greater the penalty factor. The lower.

[0068] Step S25: Adaptive Weighted Fusion. The three types of features described in Step 23 are log-normalized to stabilize the numerical scale, then weighted and summed, and multiplied by the background clutter penalty factor. To obtain the final physical saliency features ;

[0069]

[0070] in, , , These are the weighting coefficients, and .

[0071] Step S3: Use the post-processing fusion module to calculate the category confidence. and physical salience characteristics The data is then integrated to obtain a comprehensive confidence level.

[0072] In this embodiment, the post-processing fusion module adopts an adaptive fusion strategy based on exponential scaling to obtain a comprehensive confidence score, the formula of which is as follows:

[0073]

[0074] In the formula, The final overall confidence level used for non-maximum suppression (NMS) and All of these are adjustable fusion parameters.

[0075] In this embodiment, the adaptive fusion strategy has the following advantages:

[0076] when When the value is high (high physical significance of the target), the exponential term generates a gain factor greater than 1, thereby improving the final result. This makes the prediction box easier to preserve in NMS. When the value is low (the target physical significance is low), the exponential term produces a decay factor less than 1, thereby suppressing the final value. This makes the predicted box easier to filter out.

[0077] parameter control The overall strength of the feature's influence on the final confidence score; parameters Then control The shape of the response curve of the value itself. These two parameters provide a high degree of tunability for the post-training process (test application), enabling the method to be optimized according to different datasets or scenario characteristics.

[0078] Step S4: Combine the confidence level The input is the filtered bounding box data from the NMS output, which outputs the detection results and plots. Traditional NMS only... In this implementation, a tall frame is used for filtering. Selecting highly salient target boxes helps reduce the false positive rate.

[0079] Specific Implementation Method Three: Combination Figures 5 to 8 This embodiment describes an example of an infrared small target detection method that integrates physical saliency features, as described in Specific Embodiment Two:

[0080] To verify the effectiveness and superiority of the GNSD_YOLO network model described in this invention, tests were conducted on the commonly used infrared dataset IRSTD-1K, using 998 images of 512 pixels each. 512, Image example as follows Figure 5 As shown, (a) is a small target next to a cluster of urban buildings; (b) is a small target with sparse interference in the night sky; (c) is a small target in complex mountainous terrain; (d) is a small target camouflaged by cloud cover; (e) is a small target hidden by vegetation; and (f) is a small target in flat terrain.

[0081] In this embodiment, two models are used: the original YOLOv8 network and the GNSD_YOLO network model of this invention are trained in the same experimental environment using a CPU (12th Gen Intel(R) Core(TM) i5-12400F 2.50 GHz) and a GPU (NVIDIA GeForce RTX 2060). The deep learning framework used is PyTorch, and the same initial learning rate (0.001), optimizer (SGD), and training epochs (2000 epochs) are used during training to ensure fairness in the comparison. Table 1 shows the performance comparison of each model on the independent test set IRSTD-1K, which is divided into training, validation, and test sets in an 8:1:1 ratio.

[0082] Table 1

[0083]

[0084] like Figure 6 and Figure 7 As shown, Figure 6 This describes the decrease in GNSD eigenvalue loss during training. Figure 7 The loss function of the bounding box detection channel decreased. It can be seen that the loss of GNSD still maintains a good decreasing trend, indicating that adding an output channel does not affect the loss decreasing trend of that channel. The multi-scale features output by the backbone network can still be used as the output of the GNSD feature value regression channel, which demonstrates the effectiveness and feasibility of GNSD_YOLO training.

[0085] For the target detection function during training, according to Figure 6 and Figure 7 It can be observed that the detection capability was not affected during training and validation, and the training performance remained good. This indicates that the three output channels of the GNSD_YOLO detection head are relatively independent, and adding an output channel will not affect the performance of the other output channels.

[0086] The original YOLOv8 network and the trained model of GNSD_YOLO were tested on the IRSTD-1K test set. The detection results are shown in Table 2. Table 2 shows the comparative experimental results on the IRSTD-1K dataset.

[0087] Table 2

[0088]

[0089] It can be observed that, compared with the original YOLOv8 network, the GNSD_YOLO network has achieved significant improvements in core performance indicators: the mean accuracy (mAP) reached 91.92%, which is 1.32 percentage points higher than the original YOLOv8 network's 90.60%. This indicates that GNSD_YOLO has an advantage in overall detection accuracy and can more accurately locate and identify targets.

[0090] Significant reduction in false alarm rate (FPR): The model of this invention reduces the false alarm rate from 9.40% to 8.08%, a reduction of 14.0%. This significant reduction in the false alarm rate confirms that by introducing and fusing GNSD physical saliency features, the model's ability to distinguish real targets from background clutter has been substantially enhanced, effectively suppressing false alarms.

[0091] Trade-offs and Analysis of Recall: The recall rate of the model in this invention is 71.07%, slightly lower than the baseline model's 73.10%. While this invention improves accuracy and reduces false alarms during the process, it may filter out a few targets with extremely low signal-to-noise ratios in boundary cases, resulting in a slight decrease in recall. However, in applications with high real-time requirements such as early warning and guidance, reducing the false alarm rate is often more critical. Therefore, this invention has certain practical value.

[0092] In this embodiment, comparative experiments on the publicly available benchmark dataset IRSTD-1K fully validate the effectiveness of the GNSD_YOLO network model of this invention. Figure 8 As shown in the figure, the left column represents the original image, the middle column represents the GSND_YOLO prediction results, and the right column represents the YOLOv8 prediction results. The data labeled in the image represents the target detection confidence score. Based on the prediction results, it can be observed that the GNSD_YOLO network model of this invention can detect targets that YOLOv8 cannot detect, and the overall confidence score is significantly improved. This aligns with the expected result of the model's overall confidence fusion, namely, highlighting targets with high physical saliency and suppressing targets with low physical saliency. Therefore, the GNSD_YOLO network model of this invention improves the model's discrimination reliability and reduces the false alarm rate without sacrificing detection capability; furthermore, the model size and frame rate, including detection frequency, are almost identical to YOLOv8.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An infrared small target detection system that integrates physical saliency features, characterized by: The detection system includes a GNSD_YOLO network model, which consists of a feature extraction backbone network, a feature pyramid network, a target detection head, and a post-processing fusion module. The target detection head consists of three parallel bounding box regression heads, a target classification head, and a GNSD feature prediction head; The feature extraction backbone network extracts multi-scale features from the input infrared image and performs bidirectional cross-scale fusion through the feature pyramid network to output enhanced features. The enhanced features are decoded in parallel through the bounding box regression head, the target classification head, and the GNSD feature prediction head, respectively, to output the bounding box coordinates, class confidence, and physical saliency features of the target. The post-processing fusion module fuses the category confidence and physical saliency features to obtain a comprehensive confidence score, and filters the comprehensive confidence score to achieve infrared small target detection.

2. The infrared small target detection system integrating physical saliency features according to claim 1, characterized in that: The feature pyramid network consists of a bidirectional feature pyramid composed of cascaded cross-stage local network modules, bottleneck convolution modules, and upsampling modules. A cross-stage local network module is used to capture multi-scale feature contextual semantic information, and a bottleneck convolution module is used to compress channels. Then, an upsampling module is used to restore spatial resolution and output enhanced features.

3. An infrared small target detection method that integrates physical saliency features, characterized by: This detection method is implemented using an infrared small target detection system that integrates physically salient features as described in any one of claims 1 or 2; the detection method is implemented by the following steps: Step S1. Acquire a single-frame infrared image to be detected; Step S2. Input the infrared image into the pre-trained GNSD_YOLO network model; through the GNSD_YOLO network model, output the target's predicted bounding box, class confidence, and physical saliency features in parallel; Step S3. The category confidence score and physical saliency feature are fused through the post-processing fusion module to obtain a comprehensive confidence score. The comprehensive confidence score is then filtered to achieve infrared small target detection.

4. The infrared small target detection method integrating physical saliency features according to claim 3, characterized in that: In step S2, the process of obtaining the physical saliency feature is as follows: Step S21. Obtain a single-frame infrared image and a candidate bounding box, and calculate the gradient magnitude map and contrast enhancement map of the infrared image; Step S22. Extract the set of target region pixels defined by the bounding box from the infrared image, gradient magnitude map and contrast enhancement image respectively; at the same time, generate a ring-shaped region as the background region pixel set by expanding outward from the target region as the center. Step S23. Based on the pixels in the target region and the pixels in the background region, calculate the gradient contrast features, local contrast features, and region compactness features; Step S24. Calculate the background clutter penalty factor; Step S25. Adaptive weighted fusion: The gradient contrast features, local contrast features, and region compactness features obtained in step S23 are log-normalized and then weighted and summed. The sum is then multiplied by the background clutter penalty factor described in step S23 to obtain the final physical saliency features. It can be expressed as follows: ； In the formula, As a background clutter penalty factor, , , These are the weighting coefficients. Gradient contrast features, For local contrast features, It is characterized by regional compactness.

5. The infrared small target detection method integrating physical saliency features according to claim 4, characterized in that: In step S23, the gradient contrast feature is calculated. Specifically: The gradient contrast feature The mean gradient magnitude of the target region on the gradient magnitude map reflects the average gradient difference level of the target itself, and is calculated by the following formula: ； In the formula, and These are the mean and standard deviation of the gradient in the target region, respectively. and These are the mean and standard deviation of the gradient in the background region, respectively. The area of ​​the target region. To prevent division by zero, use a very small constant.

6. The infrared small target detection method integrating physical saliency features according to claim 5, characterized in that: In step S23, local contrast features are calculated. Specifically: The local contrast features The degree of contrast difference between the target area and the background area is expressed by the following formula: ； In the formula, and These are the mean values ​​of the target and background on the contrast map, respectively.

7. The infrared small target detection method integrating physical saliency features according to claim 6, characterized in that: In step S23, the regional compactness characteristics are calculated. Specifically: The regional compactness characteristics The grayscale uniformity of the target region in the infrared image is expressed by the following formula: ； in, and The mean and standard deviation of the grayscale values ​​of the target region in the infrared image.

8. The infrared small target detection method integrating physical saliency features according to claim 7, characterized in that: In step S24, the clutter level of the background region is calculated, and a penalty factor is generated. The penalty factor value is between 0 and 1; background clutter The gradient dispersion and contrast uniformity of the background region are combined, and can be expressed by the following formula: ； In the formula, The standard deviation of the background region on the contrast map; the Sigmoid function is used to... Mapped to a penalty factor: .

9. The infrared small target detection method integrating physical saliency features according to claim 3, characterized in that: The GNSD_YOLO network model, during the training phase, has an overall loss function. Based on the detection loss function and GNSD feature regression loss function The combined weighted average is obtained and expressed by the following formula: ； in, ; In the formula, To balance the hyperparameters of task weights, Set to 1, The coordinates of the center point of the target.

10. The infrared small target detection method integrating physical saliency features according to claim 3, characterized in that: In step S3, the category confidence is... Physical salience characteristics The data is then integrated to obtain a comprehensive confidence level. An adaptive fusion strategy based on exponential scaling is adopted to obtain the overall confidence score, expressed as follows: ； In the formula, To provide the final overall confidence level for nonmaximum suppression (NMS), and All of these are adjustable fusion parameters.

Images