An infrared small target detection method based on spatio-temporal context perception and compact geometric representation

Abstract

The application discloses an infrared small target detection method based on space-time context perception and compact geometric representation, comprising the following steps: acquiring three adjacent images in a sequence of infrared images to be detected; inputting a pre-trained small target detection model to output a center heat map, a center offset and an effective radius prediction result, and completing positioning and scale estimation of the infrared small target; the small target detection model is used to extract multi-time space features through a backbone network and a feature pyramid sharing weights, obtain space-time representation features suitable for infrared small target detection by introducing space-time context perception information and constructing a time domain difference enhancement and global gating adjustment mechanism, and realize direct prediction of the center position and the effective radius of the small target in combination with a decoupled geometric parameter prediction network. The infrared small target detection task is modeled as target center position and effective radius prediction, so that effective detection is realized while reducing model calculation complexity and labeling cost.

Description

Technical Field

[0001] This invention relates to the field of infrared image processing and intelligent target detection technology, specifically to an infrared small target detection method based on spatiotemporal context awareness and compact geometric representation. The method comprehensively utilizes temporal context information, motion difference information, and compact geometric representation of the target center position and effective radius in a continuous infrared image sequence to detect, locate, and estimate the scale of small targets in infrared images. Background Technology

[0002] Infrared small target detection has significant application value in scenarios such as long-range early warning, low-altitude surveillance, maritime search and rescue, and aerospace target detection. Due to the limitations of long imaging distance, weak target radiation energy, and infrared imaging mechanism, infrared small targets usually occupy only a few pixels in the image and generally have characteristics such as weak texture information, low signal-to-noise ratio, and blurred boundaries. At the same time, complex background clutter, local strong brightness interference, and imaging noise further increase the difficulty of detection.

[0003] Among existing infrared small target detection methods, a mainstream approach is to model the detection problem as a pixel-level segmentation problem and output the target region response through a neural network. While this type of method can obtain relatively fine spatial distribution results, it typically suffers from the following drawbacks: First, the foreground region of infrared small targets is extremely sparse, and the segmentation paradigm often requires the network to suppress large areas of background, resulting in a heavy computational burden, long convergence time, and hindering lightweight deployment. Second, these methods usually rely on pixel-level fine annotation, which is costly and highly subjective, limiting its application in large-scale real-world scenarios.

[0004] To reduce annotation costs and modeling complexity, some methods introduce bounding box supervision or point supervision. However, small infrared targets are affected by the diffusion effect of imaging points, typically exhibiting a compact yet blurry hotspot response. For such targets, rectangular bounding boxes are insufficient to accurately describe the effective support area of ​​the target and easily introduce significant background redundancy; while simple point representation has lower annotation costs, it only provides location information and lacks scale information, making it difficult to fully characterize the spatial support range of the target.

[0005] Furthermore, small infrared targets often appear as faint bright spots without obvious texture in a single frame image, easily lost in complex backgrounds. Some existing methods rely solely on single-frame appearance features for detection, making it difficult to fully utilize the motion continuity and temporal differences of the target between consecutive frames. This results in problems such as high false alarm rates, high false alarm rates, and unstable localization even in scenarios with weak targets, dynamic backgrounds, or complex interference.

[0006] Therefore, existing technologies still require a new infrared small target detection scheme to make fuller use of the spatiotemporal context information and compact geometric features of infrared small targets while reducing the reliance on pixel-level fine annotation, so as to achieve efficient and accurate detection, localization and scale estimation of infrared small targets. Summary of the Invention

[0007] To address the problems of existing infrared small target detection methods that rely on pixel-level segmentation modeling, have a heavy computational burden, and are heavily dependent on high-cost, fine-grained annotation, the present invention aims to provide an infrared small target detection method based on spatiotemporal context awareness and compact geometric representation.

[0008] A further objective of this invention is to address the problems of weak texture, low signal-to-noise ratio, blurred boundaries, and susceptibility to interference from complex backgrounds in single-frame images of small infrared targets by combining the temporal difference information between the two sides of the key frame and the spatial features of the current frame to improve the stable detection capability of small targets.

[0009] The present invention further involves using the target center position and effective radius to perform compact geometric modeling of small infrared targets, so as to reduce background redundancy caused by rectangular bounding box representation and realize integrated output of target localization and scale estimation.

[0010] In view of this, the present invention proposes an infrared small target detection method based on spatiotemporal context awareness and compact geometric representation, comprising: Acquire three adjacent frames from the infrared image sequence to be detected; Input a pre-established and trained small target detection model, and output the center heat map prediction result, center offset prediction result, and effective radius prediction result, thereby completing the localization and scale estimation of infrared small targets; The small target detection model is used to extract multi-time spatial features through a backbone network with shared weights and a feature pyramid. By introducing spatiotemporal context-aware information and constructing a temporal difference enhancement and global gating adjustment mechanism, it obtains spatiotemporal representation features suitable for infrared small target detection. Combined with a decoupled geometric parameter prediction network, it realizes the direct prediction of the center position and effective radius of the small target.

[0011] As an improvement to the above method, the three adjacent frames include: the previous frame image. Current frame image and the next frame image ,in For the key frame to be detected, the three adjacent frames are preprocessed to a uniform size.

[0012] As an improvement to the above method, the small target detection model includes: A shared-weight backbone network and feature pyramid are used to extract multi-temporal spatial features from three adjacent frames. , and ; The spatiotemporal context awareness module is used to combine temporal differential attention mechanism and spatial attention mechanism to obtain spatiotemporal representation features suitable for infrared small target detection; The decoupled geometric parameter prediction network includes parallel heatmap prediction heads, offset prediction heads, and radius prediction heads, which are used to output the target center heatmap, center offset, and target effective radius, respectively.

[0013] As an improvement to the above method, the shared-weight backbone network adopts a pruned lightweight convolutional neural network, uses the same set of network weights for three adjacent frames, and limits the maximum downsampling rate; the feature pyramid is used to upsample and stitch together the features of different scale levels output by the backbone network.

[0014] As an improvement to the above method, the temporal differential attention mechanism respectively addresses... and , and Perform feature difference operations, and then converge the results by addition to form a two-sided difference response, satisfying the following equation:

[0015] in, This represents the temporal difference enhancement feature at the current moment. The motion information extraction operator is represented by... It consists of convolutional layers and a sigmoid activation function. This indicates element-wise absolute value operation.

[0016] As an improvement to the above method, a motion gating adjustment unit is set within the temporal differential attention mechanism to adaptively adjust the temporal differential enhancement intensity according to the overall change in the scene; specifically including: Build global jitter strength :

[0017] in, Indicates the overall degree of change in the scene. This indicates a global average pooling operation; Based on global jitter intensity Calculate the gating weight and features of the current frame Perform temporal differential attention enhancement with motion gating to obtain the enhanced features of the current frame. :

[0018] in, This represents the learnable adjustment coefficient.

[0019] As an improvement to the above method, the processing procedure of the temporal feature fusion module includes: Features from the previous frame Enhanced current frame features and features of the next frame The features from the three frames are concatenated along the channel dimension to form a stacked feature. For stacked features Global average pooling is performed, and the pooling result is input into a multilayer perceptron to generate channel weights corresponding to the features of the previous frame, the enhanced features of the current frame, and the features of the next frame, respectively. , and Its expression is:

[0020] in, and This represents the learnable parameters in a multilayer perceptron. Represents the ReLU activation function. This represents the Sigmoid activation function. This indicates a global average pooling operation; Using channel weights , and Features of the previous frame Enhanced current frame features and features of the next frame Weighting is performed, and the features of the three weighted frames are summed and fused, then... Convolution is performed to obtain the final temporal fusion features. Its expression is:

[0021] in, express Convolution operation; A spatial attention mechanism is set after the temporal feature fusion module, and the processing includes: The spatial weight map is obtained according to the following formula. :

[0022] in, express Convolution operation, This indicates the average pooling operation. This represents the max pooling operation; Based on spatial weight graph The final detection feature map after spatial attention enhancement is obtained. : .

[0023] As an improvement to the above method, the heat map prediction head is used to output a heat map of the target center. After Sigmoid activation, a central probability heatmap is obtained. The offset prediction head is used to output a center offset map. Each candidate position corresponds to a two-dimensional offset. To compensate for the quantization error caused by downsampling; The radius prediction head is used to output an effective radius map. This characterizes the compact spatial support range of the target.

[0024] As an improvement to the above method, during the inference phase, the location of the candidate target center is determined based on the target center heatmap. And read the predicted offset value of the corresponding candidate target center position from the center offset map. Read the predicted radius value of the corresponding candidate target center position from the effective radius map. The target center coordinates at the input image scale are recovered using the following formula:

[0025] The effective radius of the target at the input image scale is then recovered using the following formula:

[0026] in, This represents the total downsampling step size of the output feature map relative to the input image. Indicates the first The center coordinates of each candidate target at the scale of the input image. Indicates the first The effective radius of each candidate target at the scale of the input image, and with... and its corresponding As the output of the detection results.

[0027] As an improvement to the above method, the method further includes training a small target detection model, wherein the supervision labels for the target center heatmap are generated using an adaptive Gaussian label assignment method; specifically including: The true effective radius of the target Based on the total downsampling step size of the output feature map relative to the input image Normalizing to the heatmap scale yields the true effective radius of the target at the heatmap scale. ; Based on the target's true effective radius at the heatmap scale Determine the Gaussian kernel standard deviation for generating Gaussian heatmap labels for the target center. , Satisfy the following formula:

[0028] in, Indicates the lower limit of standard deviation. This indicates the upper limit of the standard deviation.

[0029] As an improvement to the above method, the total loss function used in training the small object detection model is:

[0030] in, Indicates the total loss. Indicates the loss in the central heat map. This represents the center-shift regression loss. This represents the effective radius regression loss. , and These represent the weighting coefficients of the three losses.

[0031] Compared with the prior art, the advantages of the present invention are: 1. By modeling the infrared small target detection task as the prediction of the target center position and effective radius, it is possible to achieve effective detection of infrared small targets while reducing the computational complexity of the model and the labeling cost.

[0032] 2. By introducing spatiotemporal context-aware information and constructing a temporal differential enhancement and global gating adjustment mechanism, it is helpful to improve the spatiotemporal saliency of infrared small targets in complex backgrounds and enhance the detection stability of weak targets and low signal-to-noise ratio targets.

[0033] 3. By adopting a lightweight detection architecture, while ensuring high detection performance, it has low parameter quantity, computational load and inference latency, making it suitable for deployment in infrared application scenarios with requirements for real-time performance and resource consumption.

[0034] 4. By adopting compact geometric modeling of the target center position and effective radius and its corresponding decoding method, it helps to reduce background redundancy caused by rectangular bounding box representation and improve the rationality of infrared small target localization and scale estimation. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of an infrared small target detection network structure based on spatiotemporal context awareness and compact geometric representation according to the present invention. Detailed Implementation

[0036] To achieve the above objectives, this invention provides an infrared small target detection method based on spatiotemporal context awareness and compact geometric representation. This method takes three adjacent frames from a sequence of infrared images to be detected as input. First, it extracts multi-temporal spatial features using a shared-weight backbone network and a feature pyramid. Then, it utilizes a temporal differential attention mechanism, a temporal feature fusion module, and a spatial attention mechanism to obtain spatiotemporal representation features suitable for infrared small target detection. Finally, it outputs the target center heatmap, center offset, and effective target radius through a heatmap prediction head, an offset prediction head, and a radius prediction head, thereby completing the localization and scale estimation of the infrared small target.

[0037] Specifically, the method of the present invention includes at least the following steps: Step S1: Construct the input timing sequence.

[0038] Acquire multiple adjacent frames in the infrared image sequence to be detected. Preferably, select the previous frame, the current frame, and the next frame to form a three-frame input sequence, and use the current frame as the key frame to be detected.

[0039] Step S2: Extract multi-time spatial features.

[0040] The three frames of images are respectively input into a backbone network with shared weights, and the multi-scale features are upsampled and stitched together through a feature pyramid to extract the corresponding multi-time spatial features. Preferably, the maximum downsampling rate of the backbone network is limited to preserve the fine-grained spatial location information of small infrared targets.

[0041] Step S3: Perform temporal difference attention enhancement.

[0042] Based on the differences between the spatial features of the current frame and the spatial features of the preceding and following frames, a temporal differential attention enhancement feature is constructed to highlight the temporal change information related to the motion of the real target.

[0043] Step S4: Perform motion gating adjustment.

[0044] The scene change intensity is constructed based on the global differences between features of adjacent frames, and the temporal difference enhancement intensity is adaptively adjusted based on the scene change intensity to suppress pseudo-motion response caused by camera shake or overall background changes.

[0045] Step S5: Perform temporal feature fusion.

[0046] Adaptive weighted fusion is performed on the features of the previous frame, the features of the current frame after temporal enhancement, and the features of the next frame to obtain the target representation features with fused spatiotemporal context information.

[0047] Step S6: Perform spatial attention enhancement.

[0048] Spatial attention enhancement is applied to the target representation features to further suppress background clutter and highlight the suspected target region.

[0049] Step S7: Perform decoupled geometric parameter prediction.

[0050] Based on the enhanced features, the target center heatmap, center offset, and effective radius are output respectively. The geometric representation of the target uses the target center position and the target effective radius, instead of the width and height of the rectangular bounding box.

[0051] Step S8: Output the detection results.

[0052] Peak extraction is performed on the target center heatmap, preferably combined with local maximum suppression to obtain candidate center positions; then the target center coordinates are recovered by combining the center offset, and the detection result is output by combining the effective radius of the target.

[0053] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0054] Example 1 Embodiments of this invention propose an infrared small target detection method based on spatiotemporal context awareness and compact geometric representation, as shown in the appendix. Figure 1 As shown, the method mainly includes an input image sequence, a backbone network with shared weights and a feature pyramid, a temporal differential attention mechanism, a temporal feature fusion module, a spatial attention mechanism, and a heatmap prediction head, an offset prediction head, and a radius prediction head. The core of the method is to construct spatiotemporal context features on a continuous infrared image sequence and transform the infrared small target detection task into a direct prediction problem of the target center position and effective radius.

[0055] Appendix Figure 1 In the diagram, the symbol "S" inside the circle represents the feature difference operation, and the symbol "..." inside the circle represents the feature difference operation. The symbol “” indicates feature addition, the symbol “C” inside the circle indicates feature concatenation, and the cone symbol indicates upsampling.

[0056] 1. Input timing construction.

[0057] First, acquire three adjacent frames from the infrared image sequence to be detected, and denote them as the previous frame image. Current frame image and the next frame image ,in This refers to the keyframe to be detected. By introducing the temporal information before and after the keyframe, the subtle motion changes of the target in consecutive frames can be utilized to improve the temporal perception capability of weak targets, while avoiding the computational redundancy and processing delay caused by using excessively long sequences.

[0058] In this embodiment, the three infrared images are preferably preprocessed to a uniform size before being input into the network. Preferably, each infrared image can be scaled to [size missing]. Pixel.

[0059] 2. Shared backbone feature extraction.

[0060] The previous frame image Current frame image and the next frame image The corresponding multi-scale features are obtained by inputting them into a backbone network with shared weights, and then cross-scale fusion is performed through a feature pyramid to output the corresponding multi-time spatial features. , and In this context, shared weights refer to using the same network structure and sharing the same set of network weights for feature extraction across three frames of images, ensuring that the feature representations of different time frames are in a unified feature space.

[0061] In this embodiment, the backbone network preferably adopts a pruned lightweight convolutional neural network structure, and more preferably a pruned lightweight backbone network; to avoid the loss of fine-grained position information of small infrared targets during deep downsampling, it is preferable to remove excessively deep stages and limit the maximum downsampling rate to [value missing]. This is followed by the introduction of a feature pyramid network to fuse features at different scale levels. Combined with... Figure 1 The feature pyramid preferably supports hierarchical features. , and Upsampling and stitching are performed to output a high-resolution feature map suitable for subsequent temporal modeling.

[0062] The spatiotemporal context awareness module is used to combine temporal differential attention mechanism and spatial attention mechanism to obtain spatiotemporal representation features suitable for infrared small target detection.

[0063] 3. Temporal difference attention mechanism.

[0064] To enhance the salient features related to target motion in the current frame, this implementation method is based on the spatial features of the previous frame. Current frame spatial features Spatial features of the next frame Construct a temporal differential attention mechanism. Combined with the appendix... Figure 1 First, respectively and , and The characteristic difference operation is performed, and then the two-sided difference response is formed by additive convergence. The calculation formula is as follows:

[0065] in, This represents the temporal difference enhancement feature at the current moment. The motion information extraction operator is preferably derived from... It consists of convolutional layers and a sigmoid activation function, wherein the sigmoid activation function is a sigmoid activation function. This indicates element-wise absolute value operation.

[0066] 4. Motion gating adjustment.

[0067] Considering that infrared image sequences may contain camera shake, changes in viewing angle, or overall background motion, resulting in large-area pseudo-motion responses, this embodiment further incorporates a motion gating adjustment unit within the temporal differential attention mechanism. This unit adaptively adjusts the temporal differential enhancement intensity based on the overall degree of scene change. Specifically, a global shake intensity is first constructed. The calculation formula is as follows:

[0068] in, Indicates the overall degree of change in the scene. This indicates a global average pooling operation.

[0069] Then, the gating weights are calculated based on the global jitter intensity. Furthermore, it applies temporal differential attention enhancement with motion gating to the features of the current frame, the calculation formula of which is:

[0070] in, Indicates the global gating weight. This represents the learnable adjustment coefficient. This represents the features of the current frame after temporal enhancement. When the overall scene is relatively stable... Smaller, gating weight Larger network-enhanced temporal variation information; when scene jitter is severe. Increase the gating weight This reduces the likelihood of spurious motion responses, thereby automatically suppressing them.

[0071] 5. Temporal feature fusion module.

[0072] After completing motion enhancement for the current frame, this implementation further incorporates features from the previous frame. Enhanced current frame features and features of the next frame Adaptive fusion is performed to form target representation features that integrate spatiotemporal context information. Combined with appended... Figure 1 The temporal feature fusion module is located after motion gating adjustment and before the spatial attention mechanism.

[0073] Features from the previous frame Enhanced current frame features and features of the next frame The features from the three frames are concatenated along the channel dimension to form a stacked feature. For stacked features Global average pooling is performed, and the pooling result is input into a multilayer perceptron to generate channel weights corresponding to the features of the previous frame, the enhanced features of the current frame, and the features of the next frame, respectively. , and Its expression is:

[0074] in, and This represents the learnable parameters in a multilayer perceptron. This represents the ReLU activation function, which is a modified linear unit activation function; This represents the Sigmoid activation function, which is a sigmoid activation function.

[0075] The channel weight vector is split into sub-weights corresponding to the previous frame, the current frame, and the next frame, respectively. , and Then, the final temporal fusion feature is obtained through weighted fusion and convolutional smoothing, and its calculation formula is as follows:

[0076] in, This represents the temporal characteristics after fusion. express Convolution operation.

[0077] 6. Spatial attention mechanism.

[0078] After temporal feature fusion, complex background clutter and localized highlight interference may still remain in the features. To further improve the saliency of real small target regions, this embodiment describes the temporal fusion features... A spatial attention mechanism is implemented. Its spatial weight graph... The calculation formula is:

[0079] in, Represents the spatial attention weight map. express Convolution operation, This indicates the average pooling operation. This indicates a max pooling operation.

[0080] Based on the spatial weight map, the final enhanced feature map used for prediction is obtained, and its calculation formula is as follows:

[0081] in, This represents the final detection feature map after spatial attention enhancement.

[0082] 7. Decoupled geometric parameter prediction.

[0083] After obtaining the final detection feature map Subsequently, this implementation sets up three independent prediction branches, which are used to output the target center heatmap, center offset, and effective target radius, respectively. (See attached diagram.) Figure 1 The three prediction branches are the heatmap prediction head, the offset prediction head, and the radius prediction head.

[0084] The heat map prediction head is used to output the heat map of the target center. After Sigmoid activation, a center probability heatmap is obtained to represent the probability that each position in the current feature map is the center of the target; the offset prediction head is used to output the center offset map. Each candidate position corresponds to a two-dimensional offset. To compensate for the quantization error caused by downsampling; the radius prediction head is used to output the effective radius map. This is to characterize the compact spatial support range of the target. In this way, the present invention no longer uses the width and height of a rectangular bounding box to characterize small infrared targets, but instead adopts a compact geometric modeling method of "target center position + effective radius".

[0085] in, Represents the central heat map, This represents the center offset diagram. Represents the effective radius diagram. This represents the subpixel offset of the target center relative to the discrete grid position.

[0086] 8. Adaptive Gaussian label assignment.

[0087] During the training phase, let the coordinates of the true center of the target in the input image be... The total downsampling step size of the output feature map relative to the input image is Preferred First, the true center is projected onto the heatmap scale to obtain continuous center coordinates:

[0088] in, and These represent the x-coordinate and y-coordinate of the target's true center on the heatmap scale, respectively.

[0089] Then, the continuous center coordinates are mapped to the discrete grid positions to obtain the grid index where the center is located:

[0090] in, and These represent the x and y coordinates of the discrete grid corresponding to the target center, respectively. This indicates the floor function.

[0091] Based on the continuous center coordinates and discrete grid positions, construct the center offset monitoring value:

[0092] in, and These represent the horizontal and vertical offsets of the true center relative to the discrete grid position, respectively.

[0093] To ensure that the radius monitoring scale is consistent with the output scale of the heatmap, it is preferable to normalize the true effective radius of the target to the heatmap scale, resulting in:

[0094] in, This represents the true effective radius of the target at the scale of the input image. This represents the true effective radius of the target at the heatmap scale.

[0095] To enhance the monitoring adaptability to small infrared targets of different scales, this embodiment is based on the target's true effective radius at the heatmap scale. Generate adaptive Gaussian heatmap labels for the target center location, with standard deviation... The calculation formula is:

[0096] in, This represents the Gaussian kernel standard deviation. Indicates the lower limit of standard deviation. This indicates the upper limit of the standard deviation. Preferably, in On the downsampling heatmap, you can set... , .

[0097] 9. Joint loss function.

[0098] During the training phase, the total loss function is defined as:

[0099] in, Indicates the total loss. To represent the central heatmap loss, focal point loss is preferred; To represent the center-shift regression loss, the preferred method is... loss; Representing the effective radius regression loss, the preferred method is... loss; , and These represent the weighting coefficients for the three losses. Preferably, the following can be taken: , , Preferably, the offset loss and radius loss occur only at the discrete grid position corresponding to the true center of the target. Calculations are performed at that location.

[0100] 10. Output during the inference phase.

[0101] During the inference phase, the infrared image sequence to be detected is input into the trained detection network to obtain the center heatmap prediction result, center offset prediction result, and effective radius prediction result. Preferably, the heatmap prediction result is first subjected to Sigmoid activation to obtain the center probability heatmap. :

[0102] in, Represents a central probability heatmap. This represents the Sigmoid activation function, which is a sigmoid activation function.

[0103] To suppress repetitive responses at adjacent locations, local maximum suppression is preferably applied to the central probability heatmap. Preferably, this can be achieved by... Max pooling achieves heatmap-level nonmaximum suppression, and its calculation method is as follows:

[0104] in, This represents the heatmap after local maxima suppression. This represents element-wise multiplication. Represents the characteristic function, express Max pooling operation. After this process, only the peak position within the local neighborhood is retained.

[0105] Subsequently, the candidate target center positions are determined according to a threshold or sorting rule. Preferably, the candidate peak values ​​that satisfy equation (16) can be retained; in a single-target scenario, the peak position with the highest score can also be directly selected as the target center candidate.

[0106]

[0107] in, Represents the set of candidate target centers. Indicates the first Discrete grid coordinates corresponding to each candidate peak. This indicates the threshold of the heatmap.

[0108] For each candidate peak position Offset from center Read the corresponding offset prediction value From the effective radius diagram Read the corresponding radius prediction value And recover the target center position and effective target radius at the input image scale according to equations (17) and (18):

[0109] in, and They represent the first The x and y coordinates of the center of each candidate target at the scale of the input image. and These represent the corresponding predicted offsets. This represents the predicted effective radius at the heatmap scale. This represents the effective radius of the prediction restored to the scale of the input image.

[0110] In one alternative implementation, when multiple candidate peaks that are too close to each other still exist after local maximum suppression, candidate-level screening can be further performed based on candidate scores, center distance, or the overlap relationship of circular support areas to retain candidate targets with higher scores.

[0111] Combined with appendix Figure 1 The detection results are preferably displayed as the target center position and its corresponding effective radius on the current frame image. When only target localization and scale estimation are needed, the results can be directly displayed. This will be output as the final test result.

[0112] It should be noted that the training supervision labels can be derived from existing pixel-level annotation results or from other annotation extraction methods that can provide target center location and scale information; the method of generating these supervision labels does not constitute a necessary limitation on the network structure ontology of this invention. The focus of this invention is on spatiotemporal context-aware modeling, global gating adjustment, adaptive temporal fusion, spatial attention enhancement, and a network structure for predicting decoupled geometric parameters oriented towards the target center and effective radius.

[0113] The innovations of this application are mainly reflected in the following aspects: 1. A compact geometric modeling approach for small infrared targets is proposed.

[0114] In view of the imaging characteristics of infrared small targets, which are usually characterized by compact, blurry, and weakly textured thermal response regions, the target geometry is represented by the "target center position and effective radius" method instead of the traditional rectangular bounding box representation method, so that the target description is more in line with the physical imaging characteristics of infrared small targets.

[0115] 2. A spatiotemporal context-aware detection structure is proposed.

[0116] By jointly modeling the previous frame, the current frame, and the next frame, and combining temporal difference enhancement with global gating adjustment, pseudo-motion response is suppressed while highlighting the temporal changes of the real target.

[0117] 3. A lightweight, high-resolution detection architecture suitable for small infrared targets is proposed.

[0118] By employing a lightweight feature extraction network with shared parameters and limiting excessive downsampling, fine-grained spatial information of small infrared targets can be preserved while ensuring computational efficiency, thus balancing detection accuracy and engineering deployment requirements.

[0119] 4. A decoupled output scheme for geometric parameter prediction is proposed.

[0120] By predicting the target center position, center offset, and effective radius separately, the infrared small target detection task is transformed into a direct regression problem oriented towards compact geometric parameters, reducing the computational burden caused by dense segmentation modeling.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An infrared small target detection method based on spatiotemporal context awareness and compact geometric representation, comprising: Acquire three adjacent frames from the infrared image sequence to be detected; Input a pre-established and trained small target detection model, and output the center heat map prediction result, center offset prediction result, and effective radius prediction result, thereby completing the localization and scale estimation of infrared small targets; The small target detection model is used to extract multi-time spatial features through a backbone network with shared weights and a feature pyramid. By introducing spatiotemporal context-aware information and constructing a temporal difference enhancement and global gating adjustment mechanism, it obtains spatiotemporal representation features suitable for infrared small target detection. Combined with a decoupled geometric parameter prediction network, it realizes the direct prediction of the center position and effective radius of the small target.

2. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 1, characterized in that, The three adjacent frames include: the previous frame image. Current frame image and the next frame image ,in For the key frame to be detected, the three adjacent frames are preprocessed to a uniform size.

3. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 2, characterized in that, The small target detection model includes: A shared-weight backbone network and feature pyramid are used to extract multi-temporal spatial features from three adjacent frames. , and ; The spatiotemporal context awareness module is used to combine temporal differential attention mechanism and spatial attention mechanism to obtain spatiotemporal representation features suitable for infrared small target detection; The decoupled geometric parameter prediction network includes parallel heatmap prediction heads, offset prediction heads, and radius prediction heads, which are used to output the target center heatmap, center offset, and target effective radius, respectively.

4. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 3, characterized in that, The shared-weight backbone network uses a pruned, lightweight convolutional neural network, applies the same set of network weights to three adjacent frames, and limits the maximum downsampling rate; the feature pyramid is used to upsample and stitch together features of different scale levels output by the backbone network.

5. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 3, characterized in that, The temporal difference attention mechanism respectively addresses... and , and Perform feature difference operations, and then converge the results by addition to form a two-sided difference response, satisfying the following equation: in, This represents the temporal difference enhancement feature at the current moment. The motion information extraction operator is represented by... It consists of convolutional layers and a sigmoid activation function. This indicates element-wise absolute value operation.

6. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 5, characterized in that, A motion gating adjustment unit is set inside the temporal differential attention mechanism to adaptively adjust the temporal differential enhancement intensity according to the overall change of the scene; Specifically, it includes: Build global jitter strength : in, Indicates the overall degree of change in the scene. This indicates a global average pooling operation; Based on global jitter intensity Calculate the gating weight and features of the current frame Perform temporal differential attention enhancement with motion gating to obtain the enhanced features of the current frame. : in, This represents the learnable adjustment coefficient.

7. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 6, characterized in that, A temporal feature fusion module is set up within the temporal differential attention mechanism, and the processing includes: Features from the previous frame Enhanced current frame features and features of the next frame The features from the three frames are concatenated along the channel dimension to form a stacked feature. For stacked features Global average pooling is performed, and the pooling result is input into a multilayer perceptron to generate channel weights corresponding to the features of the previous frame, the enhanced features of the current frame, and the features of the next frame, respectively. , and Its expression is: in, and This represents the learnable parameters in a multilayer perceptron. Represents the ReLU activation function. This represents the Sigmoid activation function. This indicates a global average pooling operation; Using channel weights , and Features of the previous frame Enhanced current frame features and features of the next frame Weighting is performed, and the features of the three weighted frames are summed and fused, then... Convolution is performed to obtain the final temporal fusion features. Its expression is: in, express Convolution operation; A spatial attention mechanism is set after the temporal feature fusion module, and the processing includes: The spatial weight map is obtained according to the following formula. : in, express Convolution operation, This indicates the average pooling operation. This represents the max pooling operation; Based on spatial weight graph The final detection feature map after spatial attention enhancement is obtained. : 。 8. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 7, characterized in that, The heat map prediction head is used to output the heat map of the target center. After Sigmoid activation, a central probability heatmap is obtained. The offset prediction head is used to output a center offset map. Each candidate position corresponds to a two-dimensional offset. To compensate for the quantization error caused by downsampling; The radius prediction head is used to output an effective radius map. This characterizes the compact spatial support range of the target.

9. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 8, characterized in that, During the inference phase, the locations of candidate target centers are determined based on the target center heatmap. And read the predicted offset value of the corresponding candidate target center position from the center offset map. Read the predicted radius value of the corresponding candidate target center position from the effective radius map. The target center coordinates at the input image scale are recovered using the following formula: The effective radius of the target at the input image scale is then recovered using the following formula: in, This represents the total downsampling step size of the output feature map relative to the input image. Indicates the first The center coordinates of each candidate target at the scale of the input image. Indicates the first The effective radius of each candidate target at the scale of the input image, and with... and its corresponding As the output of the detection results.

10. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 3, characterized in that, The method also includes training a small target detection model, wherein the supervision labels for the target center heatmap are generated using an adaptive Gaussian label assignment method; specifically including: The true effective radius of the target Based on the total downsampling step size of the output feature map relative to the input image Normalizing to the heatmap scale yields the true effective radius of the target at the heatmap scale. ; Based on the target's true effective radius at the heatmap scale Determine the Gaussian kernel standard deviation for generating Gaussian heatmap labels for the target center. , Satisfy the following formula: in, Indicates the lower limit of standard deviation. This indicates the upper limit of the standard deviation.

11. The infrared small target detection method based on spatiotemporal context awareness and compact geometric representation according to claim 3, characterized in that, The total loss function used in training the small object detection model is: in, Indicates the total loss. Indicates the loss in the central heat map. This represents the center-shift regression loss. This represents the effective radius regression loss. , and These represent the weighting coefficients of the three losses.

Images