A marine infrared target detection method and system based on space-time feature enhancement and prediction

Abstract

The application provides a kind of based on spatiotemporal feature enhancement and prediction offshore infrared target detection method and system.The method of the application comprises: frame-by-frame feature extraction is carried out on the input infrared image sequence, reference frame is rearranged based on interframe motion similarity, the rearranged feature sequence is input into bidirectional gate recurrent unit for timing coding, and timing enhancement feature is output;The input infrared image sequence is spliced, and spatial feature enhancement is carried out through dynamic residual attention network, and spatial enhancement feature is output;The key frame of the input infrared image sequence is decomposed in multi-scale frequency domain, and the low-frequency component and the high-frequency component are enhanced respectively, and the frequency domain reconstruction and attention mechanism are refined, and the frequency domain enhancement feature is output;The timing enhancement feature, spatial enhancement feature and frequency domain enhancement feature are fused, the contribution degree of each domain feature is adjusted through adaptive weight, and the fusion feature is output;The fusion feature is input into detection head, target detection is executed, and detection result is output.

Description

Technical Field

[0001] This invention relates to the field of image processing and target detection technology, and more particularly to a method and system for detecting marine infrared targets based on spatiotemporal feature enhancement and prediction. Background Technology

[0002] With the rapid development of marine economic activities, the demand for all-weather target perception capabilities in fields such as maritime safety monitoring, fishery resource protection, and maritime search and rescue is increasing. Infrared imaging technology, due to its independence from lighting conditions, has become an important means of detecting maritime targets at night and in low-visibility environments. However, in practical applications, infrared target detection on the sea surface faces many technical challenges: the target's thermal radiation signal is weak, the image size is usually only a few pixels, and it highly overlaps with background interference such as wave clutter and solar residual heat in grayscale distribution, making it difficult for traditional detection methods to effectively distinguish the target from background noise.

[0003] Existing technologies suffer from three main shortcomings. First, they fail to fully utilize temporal information. Mainstream methods often employ independent processing of single frames or simple stacking of multiple frames, failing to fully leverage the temporal continuity of target motion, making it difficult to distinguish real targets from random noise. Second, their spatial feature extraction has limited generalization ability. Traditional convolutional networks use spatial filtering kernels with fixed parameters, making it difficult to adapt to changes in target morphology and complex backgrounds. Detection performance significantly degrades when images are blurred or targets are small. Furthermore, the application of frequency domain analysis methods is insufficient. Existing schemes fail to fully exploit frequency domain features, resulting in the ineffective utilization of the multi-scale spectral distinguishability between wave clutter and real targets. These technical deficiencies make it difficult for existing detection systems to balance detection sensitivity and false alarm suppression capabilities in complex sea conditions. A significant nonlinear trade-off exists between detection rate and false alarm rate, affecting the overall reliability of the system. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for detecting maritime infrared targets based on spatiotemporal feature enhancement and prediction. This invention proposes a detection framework that combines temporal motion analysis, adaptive spatial enhancement, and multi-scale frequency domain analysis. This framework achieves deep interaction and adaptive collaboration of information in the temporal, spatial, and frequency domains, fundamentally improving target detection performance. The method fully utilizes information from these three domains, effectively enhancing the continuity and accuracy of detection.

[0005] The technical means employed in this invention are as follows:

[0006] A method for detecting maritime infrared targets based on spatiotemporal feature enhancement and prediction, comprising: S1. Perform frame-by-frame feature extraction on the input infrared image sequence, rearrange the reference frames based on inter-frame motion similarity, input the rearranged feature sequence into a bidirectional gated recurrent unit for temporal encoding, and output temporal enhanced features. S2. The input infrared image sequence is stitched together, and spatial features are enhanced through a dynamic residual attention network to output spatial enhanced features; S3. Perform multi-scale frequency domain decomposition on the keyframes of the input infrared image sequence, enhance the low-frequency and high-frequency components respectively, and output the frequency domain enhancement features after frequency domain reconstruction and attention mechanism refinement. S4. The temporal enhancement features, spatial enhancement features, and frequency domain enhancement features are fused together, and the contribution of each domain feature is adjusted by adaptive weighting to output the fused feature. S5. Input the fused features into the detection head, perform target detection, and output the detection results.

[0007] Further, step S1 includes: S11. A deep learning-based backbone network is used to extract features from the input infrared image sequence frame by frame to obtain multi-frame feature representations. S12. Calculate the spatial centroid coordinates of the features in each frame to represent the target position, using the following formula:

[0008] in, The spatial centroid coordinates of the features in each frame are represented. This represents the average value of the feature map along the channel dimension; Represents the x and y coordinates of a pixel in the feature map; S13. Calculate the motion vector between adjacent frames based on the centroid displacement, using the following formula:

[0009] in, This represents the motion vector between adjacent frames. Indicates the first The horizontal coordinate of the spatial centroid of the frame feature map Indicates the first The spatial centroid ordinate of the frame feature map, Indicates the first The horizontal coordinate of the spatial centroid of the frame feature map Indicates the first The spatial centroid ordinate of the frame feature map; S14. Using the motion vector of the keyframe and its previous frame. Using this as a baseline, the cosine similarity of the motion vectors of each reference frame is calculated using the following formula:

[0010] in, This represents a very small constant that prevents division by zero. S15. Rearrange the reference frames in descending order of similarity, and place the reference frame with the highest similarity at the beginning to form a temporal input sequence with consistent motion trends. S16. The rearranged feature sequence is reduced in dimensionality by convolution and then input into a bidirectional gated recurrent unit. The bidirectional gated recurrent unit contains a forward GRU and a backward GRU, which encode along the forward and backward time axes respectively to capture the historical trend and future inference of the target motion. S17. Concatenate the outputs of the forward GRU and the backward GRU, and restore the original channel dimensions through convolution expansion to output a temporal enhancement feature that integrates bidirectional temporal information. .

[0011] Further, step S2 includes: S21. The input infrared image sequence is stitched together along the channel dimension to form a multi-frame aggregated feature map; S22. Input the multi-frame aggregated feature map into the dynamic residual attention network to perform spatial feature enhancement; wherein, the dynamic residual attention network includes multiple cascaded dynamic residual groups, each dynamic residual group is composed of multiple stacked dynamic residual attention blocks, and each dynamic residual attention block includes a channel attention submodule, a dynamic spatial attention submodule, and residual connections, wherein: The channel attention submodule obtains global statistical information through adaptive average pooling and adaptive max pooling, and generates channel weights through a shared multilayer perceptron; The dynamic spatial attention submodule generates dynamic convolution kernels based on input features and performs adaptive convolution operations on spatial mean features. By adding the input features of the dynamic residual attention block to the enhanced features through residual connections, gradient propagation is promoted while preserving the original feature information, resulting in enhanced features in the output space. .

[0012] Further, step S3 includes: S31. The time-series enhancement features are processed using Haar wavelet transform. Decomposed into low-frequency components Horizontal high-frequency components Vertical high frequency components and diagonal high frequency components ; S32, for low-frequency components Multi-directional local convolution enhancement is performed, which captures edge texture information through differential convolution kernels in multiple directions. The multi-directional local convolution includes differential convolution in the diagonal, vertical, anti-diagonal, and horizontal directions. S33, For horizontal high-frequency components Vertical high frequency components and diagonal high frequency components Perform direction-specific convolution enhancements to improve high-frequency details in each direction; S34. Generate adaptive weights for each frequency band component using the Sigmoid activation function. , , , The enhanced frequency band components are adaptively weighted, and the weighted frequency band components are subjected to inverse wavelet transform to reconstruct the frequency domain features. S35. The reconstructed frequency domain features are refined using a local-global attention mechanism, which includes a local attention branch and a global attention branch, wherein: The local attention branch captures local details of frequency domain features through depthwise separable convolution; The global attention branch generates global weights through adaptive average pooling and convolution operations, and adaptively weights local features; The local attention output is multiplied by the global attention output and added to the input features to output the refined frequency domain enhanced features. .

[0013] Further, in step S31, the Haar wavelet transform is implemented using a preset wavelet basis function convolution kernel, and the wavelet basis functions include: ,

[0014] ,

[0015] in, For low-frequency filters, , , These are high-frequency filters in the horizontal, vertical, and diagonal directions, respectively.

[0016] Further, step S4 includes: S41, Apply the time-enhancing features Spatial enhancement features and frequency domain enhancement features The features are concatenated along the channel dimension to form a joint feature representation of the three domains. S42. Process the joint features of the three domains through a weight generation network to output a three-channel adaptive weight map. The weight generation network includes convolutional layers and activation functions, and the weights are normalized using Softmax to satisfy... ; S43. Based on the learned adaptive weights, perform weighted fusion of features from the three domains to obtain the fused features. ,as follows:

[0017] in, The adaptive weights represent spatially enhanced features. Represents the adaptive weights of temporal enhancement features. Adaptive weights representing frequency domain enhancement features; S44, Integrating Features Channel refinement is performed using a cascaded structure of 3×3 convolutions and 1×1 convolutions to extract high-level semantic information; S45. Add the refined features and the residual features after the spatial enhancement features are transformed by a 1×1 convolution to enhance the stability of feature propagation and output the final fused features. .

[0018] Further, step S5 includes: S51, The final fusion feature Input a multi-scale detection head, and perform bounding box coordinate regression, target presence confidence prediction and target category classification through parallel convolutional branches to generate a set of candidate detection boxes; S52. Perform non-maximum suppression on candidate detection boxes, remove redundant detection boxes with an overlap exceeding a preset threshold, and retain the detection results with the highest confidence. S53. Output the final marine infrared target detection results, including target bounding box coordinates, category identifier, and confidence score.

[0019] Furthermore, the multi-scale detection head adopts a single-stage detection architecture, which directly predicts the position and category of the target on the fully convolutional feature map without the need to pre-set anchor boxes, and fuses multi-scale features through a feature pyramid structure.

[0020] This invention also provides a marine infrared target detection system based on the above-mentioned marine infrared target detection method, comprising a temporal feature enhancement module, a spatial feature enhancement module, a frequency domain feature enhancement module, a three-domain adaptive fusion module, and a detection output module, wherein: The temporal feature enhancement module is used to extract features frame by frame from the input infrared image sequence, rearrange the reference frames based on the inter-frame motion similarity, input the rearranged feature sequence into the bidirectional gated cyclic unit for temporal encoding, and output the temporal enhanced features. The spatial feature enhancement module is used to stitch together the input infrared image sequence, enhance spatial features through a dynamic residual attention network, and output spatial enhanced features. The frequency domain feature enhancement module is used to perform multi-scale frequency domain decomposition on the key frames of the input infrared image sequence, enhance the low-frequency components and high-frequency components respectively, and output frequency domain enhanced features after frequency domain reconstruction and attention mechanism refinement. The three-domain adaptive fusion module is used to fuse temporal enhancement features, spatial enhancement features and frequency domain enhancement features, and adjust the contribution of each domain feature through adaptive weights to output fused features; The detection output module is used to input the fused features into the detection head, perform target detection, and output the detection results.

[0021] Compared with the prior art, the present invention has the following advantages: 1. To address the challenges of complex target motion and difficult temporal correlation under complex sea conditions, this invention proposes a motion similarity-guided reference frame rearrangement mechanism. By calculating the similarity of motion vectors between frames, the temporal input order is optimized, making frames with consistent motion trends temporally adjacent. This significantly improves the modeling capability of the bidirectional gated loop unit for target motion trajectories and effectively solves the temporal misalignment problem caused by platform jitter and target maneuvering.

[0022] 2. To address the problem that the spatial features of infrared weak targets are weak and easily obscured by the background, this invention designs a dynamic residual attention network that combines channel attention and dynamic spatial attention mechanisms. It enhances the perception of the local structure of the target by adaptively generating spatial convolution kernels, while using residual connections to maintain the stability of feature propagation, thus significantly improving the discriminativeness of spatial feature representation.

[0023] 3. To address the problem of insufficient utilization of frequency domain information in existing methods, this invention introduces Haar wavelet transform for multi-band decomposition and designs a local-global attention mechanism to enhance frequency domain features, effectively capturing the frequency domain features of infrared targets at different scales and in different directions, and improving the algorithm's ability to suppress interference such as sea wave clutter.

[0024] 4. To address the problem that a single feature domain is insufficient to cope with complex sea conditions, this invention constructs a three-domain adaptive weighted fusion mechanism. By dynamically adjusting the contribution of spatial, temporal, and frequency domain features through a learnable weight allocation network, complementary fusion of multi-source information is achieved, significantly improving the robustness of the detection system in complex environments.

[0025] 5. This invention adopts an end-to-end deep learning architecture, which can be integrated into the existing infrared sea surface monitoring system as an independent module without modifying the original detection network structure, and has good engineering practicality and scalability.

[0026] Based on the above reasons, this invention can be widely applied in fields such as maritime safety monitoring, fishery resource protection, maritime search and rescue, waterway management, and shipborne infrared monitoring systems. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 This is a flowchart of the timing enhancement process of the present invention.

[0030] Figure 3 This is a flowchart illustrating the spatial enhancement process of the present invention.

[0031] Figure 4 This is a flowchart of the frequency domain enhancement process of the present invention.

[0032] Figure 5 This is a flowchart of the three-domain adaptive fusion and detection output process of the present invention.

[0033] Figure 6 This is a system block diagram of the present invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product or device.

[0036] like Figure 1 As shown, this invention provides a method for detecting maritime infrared targets based on spatiotemporal feature enhancement and prediction, comprising: S1. Perform frame-by-frame feature extraction on the input infrared image sequence, rearrange the reference frames based on inter-frame motion similarity, input the rearranged feature sequence into a bidirectional gated recurrent unit for temporal encoding, and output temporal enhanced features. S2. The input infrared image sequence is stitched together, and spatial features are enhanced through a dynamic residual attention network to output spatial enhanced features; S3. Perform multi-scale frequency domain decomposition on the keyframes of the input infrared image sequence, enhance the low-frequency and high-frequency components respectively, and output the frequency domain enhancement features after frequency domain reconstruction and attention mechanism refinement. S4. The temporal enhancement features, spatial enhancement features, and frequency domain enhancement features are fused together, and the contribution of each domain feature is adjusted by adaptive weighting to output the fused feature. S5. Input the fused features into the detection head, perform target detection, and output the detection results.

[0037] In a specific implementation, as a preferred embodiment of the present invention, step S1 includes: S11. A deep learning-based backbone network is used to extract features from the input infrared image sequence frame by frame to obtain multi-frame feature representations. S12. Calculate the spatial centroid coordinates of the features in each frame to represent the target position, using the following formula:

[0038] in, The spatial centroid coordinates of the features in each frame are represented. This represents the average value of the feature map along the channel dimension; Represents the x and y coordinates of a pixel in the feature map; S13. Calculate the motion vector between adjacent frames based on the centroid displacement, using the following formula:

[0039] in, This represents the motion vector between adjacent frames. Indicates the first The horizontal coordinate of the spatial centroid of the frame feature map Indicates the first The spatial centroid ordinate of the frame feature map, Indicates the first The horizontal coordinate of the spatial centroid of the frame feature map Indicates the first The spatial centroid ordinate of the frame feature map; S14. Using the motion vector of the keyframe and its previous frame. Using this as a baseline, the cosine similarity of the motion vectors of each reference frame is calculated using the following formula:

[0040] in, This represents a very small constant that prevents division by zero. S15. Rearrange the reference frames in descending order of similarity, and place the reference frame with the highest similarity at the first position to form a temporal input sequence with consistent motion trends (forming a rearranged feature sequence so that frames with consistent motion trends are temporally adjacent, optimizing the input order for subsequent temporal modeling). S16. The rearranged feature sequence is reduced in dimensionality by convolution and then input into a bidirectional gated recurrent unit. The bidirectional gated recurrent unit contains a forward GRU and a backward GRU, which encode along the forward and backward time axes respectively to capture the historical trend and future inference of the target motion. S17. Concatenate the outputs of the forward GRU and the backward GRU, and restore the original channel dimensions through convolution expansion to output a temporal enhancement feature that integrates bidirectional temporal information. In this embodiment, the number of frames The value ranges from 3 to 7 frames, preferably 5 frames; the dimension of the bidirectional GRU hidden layer is 64; and the number of layers is 2.

[0041] In this embodiment, as shown in Figure 2, the core of temporal feature enhancement lies in the motion similarity-guided reference frame reordering mechanism. This mechanism optimizes the temporal input order, enabling the bidirectional GRU to more effectively model the target motion trajectory and output temporal enhanced features.

[0042] In a specific implementation, as a preferred embodiment of the present invention, as shown in Figure 3, step S2 includes: S21. The input infrared image sequence is stitched together along the channel dimension to form a multi-frame aggregated feature map; S22. Input the multi-frame aggregated feature map into the dynamic residual attention network to perform spatial feature enhancement; wherein, the dynamic residual attention network includes multiple cascaded dynamic residual groups, each dynamic residual group is composed of multiple stacked dynamic residual attention blocks, and each dynamic residual attention block includes a channel attention submodule, a dynamic spatial attention submodule, and residual connections, wherein: The channel attention submodule obtains global statistical information through adaptive average pooling and adaptive max pooling, and generates channel weights through a shared multilayer perceptron; The dynamic spatial attention submodule generates dynamic convolutional kernels based on input features and performs adaptive convolution operations on spatial mean features. In this embodiment, dynamic convolutional kernel parameters are generated through global average pooling and convolution operations. The size of the dynamic convolution kernel is Preferably, the value is 3×3. The mean of the input features along the channel dimension is... With dynamic convolution kernels Perform depthwise separable convolution operations to generate a spatial attention map, activate it with a sigmoid function, multiply it with the input features, and output dynamic spatial attention-enhanced features. By adding the input features of the dynamic residual attention block to the enhanced features through residual connections, gradient propagation is promoted while preserving the original feature information, resulting in enhanced features in the output space. .

[0043] In this embodiment, the number of dynamic residual groups ranges from 3 to 6, preferably 5; the number of dynamic residual attention blocks within each dynamic residual group ranges from 3 to 6, preferably 5. The core advantage of the dynamic residual attention network lies in its dynamic spatial attention mechanism. Unlike traditional static convolutional kernels, the generation process of dynamic convolutional kernels is related to the input feature content, enabling adaptive adjustment of filtering characteristics for different spatial locations, effectively enhancing the feature saliency of small infrared targets in complex backgrounds.

[0044] In specific implementation, as a preferred embodiment of the present invention, such as Figure 4 As shown, step S3 includes: S31. The time-series enhancement features are processed using Haar wavelet transform. Decomposed into low-frequency components Horizontal high-frequency components Vertical high frequency components and diagonal high frequency components ; S32, for low-frequency components Multi-directional local convolution enhancement is performed, which captures edge texture information through differential convolution kernels in multiple directions. The multi-directional local convolution includes differential convolution in the diagonal, vertical, anti-diagonal, and horizontal directions. S33, For horizontal high-frequency components Vertical high frequency components and diagonal high frequency components Perform direction-specific convolution enhancements to improve high-frequency details in each direction; S34. Generate adaptive weights for each frequency band component using the Sigmoid activation function. , , , The enhanced frequency band components are adaptively weighted, and the weighted frequency band components are subjected to inverse wavelet transform to reconstruct the frequency domain features. S35. The reconstructed frequency domain features are refined using a local-global attention mechanism, which includes a local attention branch and a global attention branch, wherein: The local attention branch captures local details of frequency domain features through depthwise separable convolution; The global attention branch generates global weights through adaptive average pooling and convolution operations, and adaptively weights local features; The local attention output is multiplied by the global attention output and added to the input features to output the refined frequency domain enhanced features. .

[0045] In this embodiment, Haar wavelet transform can effectively separate the energy distribution of the target and background in different frequency bands. Low-frequency components retain the overall structural information of the target, while high-frequency components capture edge and detail information. Through frequency band adaptive enhancement, the energy response of background interference such as ocean clutter in specific frequency bands can be effectively suppressed, improving the distinguishability between the target and the background.

[0046] In a specific implementation, as a preferred embodiment of the present invention, in step S31, the Haar wavelet transform is implemented through a preset wavelet basis function convolution kernel, and the wavelet basis functions include: ,

[0047] ,

[0048] in, For low-frequency filters, , , These are high-frequency filters in the horizontal, vertical, and diagonal directions, respectively.

[0049] In specific implementation, as a preferred embodiment of the present invention, such as Figure 5 As shown, step S4 includes: S41, Apply the time-enhancing features Spatial enhancement features and frequency domain enhancement features The features are concatenated along the channel dimension to form a joint feature representation of the three domains. S42. Process the joint features of the three domains through a weight generation network to output a three-channel adaptive weight map. The weight generation network includes convolutional layers and activation functions, and the weights are normalized using Softmax to satisfy... ; S43. Based on the learned adaptive weights, perform weighted fusion of features from the three domains to obtain the fused features. ,as follows:

[0050] in, The adaptive weights represent spatially enhanced features. Represents the adaptive weights of temporal enhancement features. Adaptive weights representing frequency domain enhancement features; S44, Integrating Features Channel refinement is performed using a cascaded structure of 3×3 convolutions and 1×1 convolutions to extract high-level semantic information; S45. Add the refined features and the residual features after the spatial enhancement features are transformed by a 1×1 convolution to enhance the stability of feature propagation and output the final fused features. .

[0051] In this embodiment, the core of the three-domain adaptive weighted fusion mechanism lies in the fact that the weight generation network can dynamically adjust the contribution of the three-domain features according to the content of the current input features. When the target motion is obvious, the weight of temporal features is enhanced; when the target structure is clear, the weight of spatial features is enhanced; and when the background interference is complex, the weight of frequency domain features is enhanced, thereby achieving adaptive feature selection and fusion.

[0052] In a specific implementation, as a preferred embodiment of the present invention, step S5 includes: S51, The final fusion feature Input a multi-scale detection head, and perform bounding box coordinate regression, target presence confidence prediction and target category classification through parallel convolutional branches to generate a set of candidate detection boxes; S52. Perform non-maximum suppression on candidate detection boxes, remove redundant detection boxes with an overlap exceeding a preset threshold, and retain the detection results with the highest confidence. S53. Output the final marine infrared target detection results, including target bounding box coordinates, category identifier, and confidence score.

[0053] In a specific implementation, as a preferred embodiment of the present invention, the multi-scale detection head adopts a single-stage detection architecture, which directly predicts the position and category of the target on the fully convolutional feature map without the need to pre-set anchor boxes. It improves the detection capability of targets at different scales by fusing multi-scale features through a feature pyramid structure.

[0054] like Figure 6 As shown, the present invention also provides a marine infrared target detection system based on the above-mentioned marine infrared target detection method, including a temporal feature enhancement module, a spatial feature enhancement module, a frequency domain feature enhancement module, a three-domain adaptive fusion module, and a detection output module, wherein: The temporal feature enhancement module is used to extract features frame by frame from the input infrared image sequence, rearrange the reference frames based on the inter-frame motion similarity, input the rearranged feature sequence into the bidirectional gated cyclic unit for temporal encoding, and output the temporal enhanced features. The spatial feature enhancement module is used to stitch together the input infrared image sequence, enhance spatial features through a dynamic residual attention network, and output spatial enhanced features. The frequency domain feature enhancement module is used to perform multi-scale frequency domain decomposition on the key frames of the input infrared image sequence, enhance the low-frequency components and high-frequency components respectively, and output frequency domain enhanced features after frequency domain reconstruction and attention mechanism refinement. The three-domain adaptive fusion module is used to fuse temporal enhancement features, spatial enhancement features and frequency domain enhancement features, and adjust the contribution of each domain feature through adaptive weights to output fused features; The detection output module is used to input the fused features into the detection head, perform target detection, and output the detection results.

[0055] The embodiments of the present invention are described simply because they correspond to those in the embodiments above. For any similarities, please refer to the descriptions in the embodiments above, which will not be elaborated here.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting maritime infrared targets based on spatiotemporal feature enhancement and prediction, characterized in that, include: S1. Perform frame-by-frame feature extraction on the input infrared image sequence, rearrange the reference frames based on inter-frame motion similarity, input the rearranged feature sequence into a bidirectional gated recurrent unit for temporal encoding, and output temporal enhanced features. S2. The input infrared image sequence is stitched together, and spatial features are enhanced through a dynamic residual attention network to output spatial enhanced features; S3. Perform multi-scale frequency domain decomposition on the keyframes of the input infrared image sequence, enhance the low-frequency and high-frequency components respectively, and output the frequency domain enhancement features after frequency domain reconstruction and attention mechanism refinement. S4. The temporal enhancement features, spatial enhancement features, and frequency domain enhancement features are fused together, and the contribution of each domain feature is adjusted by adaptive weighting to output the fused feature. S5. Input the fused features into the detection head, perform target detection, and output the detection results.

2. The maritime infrared target detection method based on spatiotemporal feature enhancement and prediction according to claim 1, characterized in that, Step S1 includes: S11. A deep learning-based backbone network is used to extract features from the input infrared image sequence frame by frame to obtain multi-frame feature representations. S12. Calculate the spatial centroid coordinates of the features in each frame to represent the target position, using the following formula: in, The spatial centroid coordinates of the features in each frame are represented. This represents the average value of the feature map along the channel dimension; Represents the x and y coordinates of a pixel in the feature map; S13. Calculate the motion vector between adjacent frames based on the centroid displacement, using the following formula: in, This represents the motion vector between adjacent frames. Indicates the first The horizontal coordinate of the spatial centroid of the frame feature map Indicates the first The spatial centroid ordinate of the frame feature map, Indicates the first The horizontal coordinate of the spatial centroid of the frame feature map Indicates the first The spatial centroid ordinate of the frame feature map; S14. Using the motion vector of the keyframe and its previous frame. Using this as a baseline, the cosine similarity of the motion vectors of each reference frame is calculated using the following formula: in, This represents a very small constant that prevents division by zero. S15. Rearrange the reference frames in descending order of similarity, and place the reference frame with the highest similarity at the beginning to form a temporal input sequence with consistent motion trends. S16. The rearranged feature sequence is reduced in dimensionality by convolution and then input into a bidirectional gated recurrent unit. The bidirectional gated recurrent unit contains a forward GRU and a backward GRU, which encode along the forward and backward time axes respectively to capture the historical trend and future inference of the target motion. S17. Concatenate the outputs of the forward GRU and the backward GRU, and restore the original channel dimensions through convolution expansion to output a temporal enhancement feature that integrates bidirectional temporal information. .

3. The maritime infrared target detection method based on spatiotemporal feature enhancement and prediction according to claim 1, characterized in that, Step S2 includes: S21. The input infrared image sequence is stitched together along the channel dimension to form a multi-frame aggregated feature map; S22. Input the multi-frame aggregated feature map into the dynamic residual attention network to perform spatial feature enhancement; wherein, the dynamic residual attention network includes multiple cascaded dynamic residual groups, each dynamic residual group is composed of multiple stacked dynamic residual attention blocks, and each dynamic residual attention block includes a channel attention submodule, a dynamic spatial attention submodule, and residual connections, wherein: The channel attention submodule obtains global statistical information through adaptive average pooling and adaptive max pooling, and generates channel weights through a shared multilayer perceptron; The dynamic spatial attention submodule generates dynamic convolution kernels based on input features and performs adaptive convolution operations on spatial mean features. By adding the input features of the dynamic residual attention block to the enhanced features through residual connections, gradient propagation is promoted while preserving the original feature information, resulting in enhanced features in the output space. .

4. The maritime infrared target detection method based on spatiotemporal feature enhancement and prediction according to claim 1, characterized in that, Step S3 includes: S31. The time-series enhancement features are processed using Haar wavelet transform. Decomposed into low-frequency components Horizontal high-frequency components Vertical high frequency components and diagonal high frequency components ; S32, for low-frequency components Multi-directional local convolution enhancement is performed, which captures edge texture information through differential convolution kernels in multiple directions. The multi-directional local convolution includes differential convolution in the diagonal, vertical, anti-diagonal, and horizontal directions. S33, For horizontal high-frequency components Vertical high frequency components and diagonal high frequency components Perform direction-specific convolution enhancements to improve high-frequency details in each direction; S34. Generate adaptive weights for each frequency band component using the Sigmoid activation function. , , , The enhanced frequency band components are adaptively weighted, and the weighted frequency band components are subjected to inverse wavelet transform to reconstruct the frequency domain features. S35. The reconstructed frequency domain features are refined using a local-global attention mechanism, which includes a local attention branch and a global attention branch, wherein: The local attention branch captures local details of frequency domain features through depthwise separable convolution; The global attention branch generates global weights through adaptive average pooling and convolution operations, and adaptively weights local features; The local attention output is multiplied by the global attention output and added to the input features to output the refined frequency domain enhanced features. .

5. The method for detecting maritime infrared targets based on spatiotemporal feature enhancement and prediction according to claim 4, characterized in that, In step S31, the Haar wavelet transform is implemented using a preset wavelet basis function convolution kernel, and the wavelet basis functions include: , , in, For low-frequency filters, , , These are high-frequency filters in the horizontal, vertical, and diagonal directions, respectively.

6. The method for detecting maritime infrared targets based on spatiotemporal feature enhancement and prediction according to claim 1, characterized in that, Step S4 includes: S41, Apply the time-enhancing features Spatial enhancement features and frequency domain enhancement features The features are concatenated along the channel dimension to form a joint feature representation of the three domains. S42. Process the joint features of the three domains through a weight generation network to output a three-channel adaptive weight map. The weight generation network includes convolutional layers and activation functions, and the weights are normalized using Softmax to satisfy... ; S43. Based on the learned adaptive weights, perform weighted fusion of features from the three domains to obtain the fused features. ,as follows: in, The adaptive weights represent spatially enhanced features. Represents the adaptive weights of temporal enhancement features. Adaptive weights representing frequency domain enhancement features; S44, Integrating Features Channel refinement is performed using a cascaded structure of 3×3 convolutions and 1×1 convolutions to extract high-level semantic information; S45. Add the refined features and the residual features after the spatial enhancement features are transformed by a 1×1 convolution to enhance the stability of feature propagation and output the final fused features. .

7. The maritime infrared target detection method based on spatiotemporal feature enhancement and prediction according to claim 1, characterized in that, Step S5 includes: S51, The final fusion feature Input a multi-scale detection head, and perform bounding box coordinate regression, target presence confidence prediction and target category classification through parallel convolutional branches to generate a set of candidate detection boxes; S52. Perform non-maximum suppression on candidate detection boxes, remove redundant detection boxes with an overlap exceeding a preset threshold, and retain the detection results with the highest confidence. S53. Output the final marine infrared target detection results, including target bounding box coordinates, category identifier, and confidence score.

8. The method for detecting maritime infrared targets based on spatiotemporal feature enhancement and prediction according to claim 7, characterized in that, The multi-scale detection head adopts a single-stage detection architecture, which directly predicts the position and category of the target on the fully convolutional feature map without the need to pre-set anchor boxes, and fuses multi-scale features through a feature pyramid structure.

9. A marine infrared target detection system based on the marine infrared target detection method according to any one of claims 1-8, characterized in that, It includes a temporal feature enhancement module, a spatial feature enhancement module, a frequency domain feature enhancement module, a three-domain adaptive fusion module, and a detection output module, wherein: The temporal feature enhancement module is used to extract features frame by frame from the input infrared image sequence, rearrange the reference frames based on the inter-frame motion similarity, input the rearranged feature sequence into the bidirectional gated cyclic unit for temporal encoding, and output the temporal enhanced features. The spatial feature enhancement module is used to stitch together the input infrared image sequence, enhance spatial features through a dynamic residual attention network, and output spatial enhanced features. The frequency domain feature enhancement module is used to perform multi-scale frequency domain decomposition on the key frames of the input infrared image sequence, enhance the low-frequency components and high-frequency components respectively, and output frequency domain enhanced features after frequency domain reconstruction and attention mechanism refinement. The three-domain adaptive fusion module is used to fuse temporal enhancement features, spatial enhancement features and frequency domain enhancement features, and adjust the contribution of each domain feature through adaptive weights to output fused features; The detection output module is used to input the fused features into the detection head, perform target detection, and output the detection results.

Images