Video infrared image denoising enhancement method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-03-27
Smart Images
Figure CN121746231A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of infrared image processing technology, specifically relating to a method for denoising and enhancing video infrared images. Background Technology
[0002] With the widespread application of infrared imaging systems in various scenarios, the requirements for image quality are constantly increasing. Especially when airborne platforms such as helicopters and drones perform flight photography missions, infrared images are easily affected by the following factors: (1) The high-speed motion of the platform causes significant displacement or blurring between image frames; (2) Atmospheric disturbances and heat flow fluctuations cause bias fields and strip noise; (3) Gaussian thermal noise interference inherent in the infrared sensor itself; (4) The imaging system has limited modulation capability, resulting in insufficient target contrast and blurred details. Summary of the Invention
[0003] Purpose of the invention: To address the aforementioned technical problems, this invention provides an infrared image quality improvement method for aviation platforms that combines video optical flow estimation with multi-frame fusion denoising using a Transformer structure and incorporates a nonlinear histogram enhancement mechanism. This method is applicable to enhancing the temporal consistency and detail of infrared imaging equipment mounted on helicopters in air-to-ground and air-to-air missions.
[0004] This application provides a method for denoising and enhancing infrared video images, the method comprising: Step 1: Perform grayscale normalization and size adjustment on the acquired continuous infrared image sequence to obtain a standardized image sequence. ; Step 2: Construct a pyramid-shaped framework for the continuous infrared image sequence after grayscale normalization and resizing. Process it; Step 3: Based on the pyramid-shaped framework, extract the continuous infrared image sequence. The optical flow characteristics; wherein, the optical flow characteristics include a forward optical flow field. and reverse optical flow field ; Step 4: Construct a hierarchical denoising network for continuous infrared image sequences. Denoising is performed on mid-infrared images to obtain feature sequences; Step 5: Based on the forward optical flow field and reverse optical flow field The feature sequence is enhanced to obtain the enhanced feature sequence; Step 6: Based on the enhanced feature sequence, a dynamic nonlinear grayscale enhancement module based on histogram modeling is introduced to improve the contrast and structural clarity of the infrared image after noise removal, and enhance the visual perception of thermal targets in the image.
[0005] Preferably, step 1 includes: Acquiring continuous infrared image sequences And perform grayscale normalization and size adjustment. This adapts to subsequent network input; where H is the image height and W is the image width.
[0006] Preferably, step 2 includes: a) Input frame and The pieces are spliced along the channel dimension to form a size of Input tensor ; b) In the first layer, for the input tensor By two consecutive times After convolution and GELU activation, preliminary edge and texture information is extracted to obtain a 64-channel feature map. Subsequently, max pooling is used for spatial downsampling, reducing the size to half of the original, i.e., the image height is H / 2 and the image width is W / 2. c) The second layer uses the same double convolutional structure as before to extract higher-level contextual information, and further reduces the spatial resolution through max pooling. The image height becomes H / 4, the image width becomes W / 4, and the feature dimension is increased to 128 channels. d) The third layer continues to apply the double residual structure downsampling, the image height becomes H / 8, the image width becomes W / 8, and a 256-channel feature map is obtained, which has the ability to represent the deformation of objects and the global structure. e) The fourth layer is the deepest layer. The image height becomes H / 16 and the image width becomes W / 16. Double convolution is used to increase the feature dimension to 512 channels while preserving the global semantic features with the lowest spatial resolution. f) The feature map of the fifth layer is first subjected to bilinear upsampling to restore the size to twice that of the previous layer, the image height becomes H / 8, the image width becomes W / 8, and the channel dimension is reduced to 256 through a convolution operation; then it is concatenated with the feature map of the third layer symmetrical to the encoder to fuse high-level semantics and low-level detail information. g) Repeat the same operations for each subsequent level: upsampling, convolution, and concatenation; the corresponding fused layers are the second layer and the first layer, with convolution kernels of all... The output channels are 128 and 64 respectively; h) In the last layer, the high-resolution feature map obtained by the decoder is mapped to a two-dimensional optical flow field with a size of , whose two channels respectively correspond to the horizontal and vertical motion of each pixel in the image.
[0007] Preferably, the step 3 comprises: a) inputting the forward optical flow field into the above framework to obtain the inverse optical flow field . b) Then input the inverse optical flow field into the above framework to obtain the inverse optical flow field .
[0008] Preferably, the step 4 comprises: a) inputting the infrared image sequence into a convolutional encoder for feature extension, which is composed of multiple two-dimensional convolutional layers, and outputting a feature tensor:
[0009] wherein, is the feature encoder. This step obtains a time-series feature sequence with feature extension:
[0010] wherein, is the total number of feature channels, which is divided into three sub-channel paths according to the modeling requirements of different scales; b) Next, the feature tensor is split into three groups of sub-channel features according to the channel dimension, and is respectively denoted as:
[0011]
[0012]
[0013] Each path of features first passes through a convolutional kernel to complete channel compression and dimension alignment for consistent input for subsequent attention calculation; c) The above three features are respectively input into three time-series mutual attention modules; each module includes a multi-head self-attention calculation structure and introduces different time-series down-sampling rate parameters ρ: The first branch uses ρ=1, indicating that all frames participate in attention modeling; The second branch uses ρ=2, and time-series sampling is performed every 2 frames; The third branch uses ρ=3, and sampling is performed every 3 frames, focusing on long-time dependency modeling; d) In a pair of images, the reference frame and the support frame, and they are the reference frame and the support frame, respectively, respectively , the reference frame , its features are converted into query vectors through a set of learnable linear mapping weights :
[0014] For each frame in the support frame , the key and the value are generated respectively through shared or independent weights:
[0015] where, are the mapping matrices of the key and the value respectively, and d is the internal dimension set in the attention calculation; e) Any frame The information aggregation form in the attention mechanism is as follows:
[0016] where, represents the support frame set selected according to the down-sampling rate p, , is the attention weight, which is calculated from the reference frame and the support frame ;
[0017] All attention operations are learnable and optimized end-to-end in training; f) Each attention output feature is input into a set of feed-forward neural network modules (FFN) for information enhancement, and the structure is usually as follows: the first layer: convolution + GELU activation; the second layer: convolution + residual connection, which can be expressed by the formula:
[0018] g) The enhanced features output by the three branches are respectively, which are spliced in the channel dimension:
[0019] where, The first step is to perform a dimensional concatenation operation, followed by feature compression and fusion using a 1×1 convolutional kernel to restore the channel dimension to 1, resulting in the final output. .
[0020] Preferably, step 5 includes: a) For the feature sequence output by the temporal mutual attention module Furthermore, an enhancement mechanism based on optical flow guidance is introduced to improve the consistency and continuity of cross-frame features, especially to suppress detail blurring or target drift caused by factors such as attention mismatch and inter-frame deformation. b) Utilizing the forward optical flow field from step 3 and reverse optical flow field The output features of the preceding and following frames are transformed to the coordinate system of the current frame, respectively, to obtain the forward enhancement features and the backward enhancement features:
[0021]
[0022] in, This represents a bilinear interpolation operation based on optical flow, used to map features from different frames to their spatiotemporal locations in a reference frame. These are forward enhancement features and backward enhancement features, respectively; c) Output features of the current frame The features of adjacent frames aligned with it are fused to form a final feature representation with stronger structural consistency:
[0023] d) Compress the final feature representation through a 1×1 convolution:
[0024] e) Finally, the enhanced feature sequence is obtained. : .
[0025] 7. The method according to claim 1, wherein step 6 comprises: a) First, calculate the grayscale distribution of the global or local regions of the image. Construct a normalized histogram:
[0026] in, For the Kronecker function, ∈[0,255] represents integer gray levels. Based on this distribution, a cumulative distribution function (CDF) is then constructed:
[0027] b) To address the differences in brightness distribution across different images, an adaptive gamma factor γ is introduced to perform a nonlinear transformation on the CDF:
[0028] in, For adjustment coefficients, Represents image entropy; c) Divide the image into several M×M non-overlapping blocks, independently calculate the histogram and CDF transform function of each block, and then perform enhancement:
[0029] in, The modulated mapping function, It is the fusion weight of each local block for pixel (x,y); d) To prevent over-enhancement and artifacts near the target edges, an edge gradient-aware control mechanism is introduced:
[0030] For pixel locations where the gradient value is greater than a set threshold τ, reduce the enhancement amplitude: .
[0031] Preferably, edge adaptive weights Defined as:
[0032] β controls the edge suppression intensity, ensuring edge continuity and structural integrity, ultimately resulting in an enhanced infrared image. .
[0033] This application has the following technical advantages: 1) Significantly improves image quality in complex flight environments To address the multi-source interference problems such as motion blur, thermal distortion, and offset striping that exist when helicopter platforms perform air-to-ground or air-to-air reconnaissance missions, this invention can effectively suppress noise, preserve the structural details of thermal targets, and achieve high signal-to-noise ratio image restoration.
[0034] 2) Enhance inter-frame temporal consistency to facilitate subsequent target identification. By employing a mutual attention mechanism that combines optical flow estimation and inter-frame soft alignment, the problem of inter-frame feature drift is solved, ensuring that the contour and grayscale of the thermal target remain continuous in consecutive frames, thus providing stable input for subsequent detection, recognition, and tracking tasks.
[0035] 3) Combines explicit and implicit alignment methods, providing strong modeling capabilities. This invention combines explicit alignment based on optical flow with implicit modeling of Transformer to achieve complementary enhancement effects of inter-frame alignment and structural reinforcement, which is particularly suitable for low signal-to-noise infrared scenes.
[0036] 4) Enhance target perceptibility and contrast The dynamic nonlinear enhancement module adjusts the stretching amplitude according to the statistical characteristics of the image, effectively enhancing the visual response of weak thermal target areas, and is suitable for weak target scenarios such as night vision and remote observation.
[0037] 5) Strong edge stability, suppressing artifact occurrence After introducing the gradient sensing mechanism, the enhancement effect has a natural transition capability in the edge region, which can prevent structural breakage and halo artifacts and is beneficial to the visual consistency of infrared images.
[0038] 6) Possesses end-to-end deployment capabilities and embedded system compatibility Each module employs a lightweight convolutional structure and a separable attention mechanism, making it suitable for deployment on platforms with limited computing resources, such as drones and airborne equipment, and highly practical. Attached Figure Description
[0039] Figure 1 This is a schematic diagram of the pyramid optical flow calculation framework; Figure 2 This is a schematic diagram of a hierarchical denoising network; Figure 3 This is a schematic diagram of the inter-frame attention mechanism; Figure 4 This is a schematic diagram of inter-frame enhancement based on optical flow; Figure 5 This is a schematic diagram illustrating the effect of the present invention. Detailed Implementation This application provides an infrared image quality improvement method for aviation platforms that combines video optical flow estimation with Transformer structure for multi-frame fusion denoising and integrates a nonlinear histogram enhancement mechanism. It is applicable to the imaging temporal consistency enhancement and detail enhancement of helicopter-borne infrared imaging equipment in air-to-ground and air-to-air missions.
[0040] Please see Figures 1-5 The solution provided in this application specifically includes the following steps: Step 1: Acquire a continuous infrared image sequence, perform grayscale normalization, and resize to 512*512 to obtain the input sequence. ; Step 2, as follows Figure 1 As shown, a pyramid-shaped framework is constructed for extracting optical flow features from infrared image sequences; a) Take two adjacent frames of the input sequence and The pieces are spliced along the channel dimension to form a size of The input tensor; b) In the first layer, through two consecutive... After convolution and GELU activation, preliminary edge and texture information is extracted to obtain a 64-channel feature map. Subsequently, max pooling is used for spatial downsampling, reducing the size to half of the original, i.e., the image height is 256 and the image width is 256. c) The second layer uses the same double convolutional structure as before to extract higher-level contextual information, and further reduces the spatial resolution through max pooling. The image height becomes 128, the image width becomes 128, and the feature dimension is increased to 128 channels. d) The third layer continues to apply a double convolutional structure and downsamples, that is, the image height becomes 64 and the image width becomes 64, obtaining a 256-channel feature map, which has the ability to represent the shape of objects and the global structure. e) The fourth layer is the deepest layer. The image height and the image width become 32. Double convolution is used to increase the feature dimension to 512 channels, while preserving global semantic features with extremely low spatial resolution. f) The feature map of the fifth layer is first subjected to bilinear upsampling, which enlarges the size to twice that of the previous layer, making the image height 64 and the image width 64. Then, a convolution operation is used to compress the number of channels to 256. Subsequently, it is concatenated with the feature map of the third layer at the symmetrical position of the encoder to fuse high-level semantics and low-level detail information. g) Repeat the same operation for each subsequent level: upsampling, convolution, and concatenation, corresponding to the fusion of the encoder feature maps from the second and first layers. Each convolutional layer uses a 3×3 kernel, with output channels of 128 and 64 respectively. h) In the last layer, the high-resolution feature map obtained by the decoder is mapped to a two-dimensional optical flow field of size 512×512×2 through an output convolutional layer, with its two channels corresponding to the horizontal and vertical motion of each pixel in the image, respectively.
[0041] Step 3, Inputting the above framework yields the forward optical flow field. Then put Inputting the above framework yields the reverse optical flow field. ; Step 4, as follows Figure 2 As shown, a hierarchical denoising network is constructed to denoise infrared images; a) Infrared image sequence The input is fed into a convolutional encoder for feature stretching. This encoder consists of multiple 2D convolutional layers, including activation functions and normalization modules, and its output is a feature tensor.
[0042] in, This is the feature encoder. This step yields the temporal feature sequence for feature expansion:
[0043] in, The total number of feature channels is pre-divided into three sub-channels according to the modeling requirements at different scales; b) Next, this feature tensor The features are split into three sub-channels based on the channel dimension, denoted as follows:
[0044]
[0045]
[0046] Each feature first passes through a The convolutional kernel performs channel compression and dimension alignment, serving as a consistent input for subsequent attention calculations. c) The three features mentioned above are respectively input into three temporal mutual attention modules; each module contains a multi-head self-attention computation structure and introduces different temporal downsampling rate parameters ρ: The first branch uses ρ=1, indicating that the entire frame participates in attention modeling; The second branch uses ρ=2, and performs temporal sampling every 2 frames (1 frame apart); The third branch uses ρ=3 and samples every 3 frames (2 frames apart) to focus on long-term dependency modeling.
[0047] d) such as Figure 3 As shown, in a pair of images, the reference frame and the supporting frame are mutually referenced. For example... and They serve as each other's reference and support frames. As a reference frame, To support frames For the reference frame Its features are passed through a set of learnable linear mapping weights. Convert to query vector :
[0048] For each frame in the support frames Keys are generated using either shared or independent weights. AND value :
[0049] in, Here, represents the mapping matrix between keys and values, and d is the internal dimension set in the attention calculation. e) Using any frame For example, its information aggregation form in the attention mechanism is as follows:
[0050] in, This represents the set of support frames selected based on the downsampling rate ρ. , Attention weights are determined by the reference frame. With supporting frames Calculated;
[0051] All attention operations are learnable and optimized end-to-end during training.
[0052] f) Features of each attention output Each input is a set of feedforward neural network (FFN) modules for information augmentation. The structure is typically: First layer: convolution + GELU activation; Second layer: convolution + residual connection. This can be expressed as:
[0053] g) The enhanced features of the three branches are as follows: They are spliced together along the channel dimension:
[0054] in This involves dimensional concatenation. Then, a 1×1 convolutional kernel is used for feature compression and fusion, restoring the channel dimension to its initial 1, resulting in the final output.
[0055] Step 5: Enhance output features based on optical flow; a) For the feature sequence output by the temporal mutual attention module Furthermore, an enhancement mechanism based on optical flow guidance is introduced to improve the consistency and continuity of cross-frame features, especially to suppress detail blurring or target drift caused by factors such as attention mismatch and inter-frame deformation. b) Utilizing the forward optical flow field from step 3 and reverse optical flow field The output features of the preceding and following frames are transformed to the coordinate system of the current frame, respectively. This yields the forward enhancement features and the backward enhancement features:
[0056]
[0057] in This represents a bilinear interpolation operation based on optical flow, used to map features from different frames to the spatiotemporal location of a reference frame. These are forward enhancement features and backward enhancement features, respectively; c) Output features of the current frame The features of adjacent frames aligned with it are fused to form a final feature representation with stronger structural consistency:
[0058] d) Compress the final feature representation through a 1×1 convolution:
[0059] e) Finally, the enhanced feature sequence is obtained. :
[0060] Step 6: Further, in order to improve the contrast and structural clarity of the infrared image after noise removal, and enhance the visual perception of thermal targets in the image, the enhanced image feature sequence... Based on this, a dynamic nonlinear grayscale enhancement module based on histogram modeling is introduced, and its processing flow is as follows: a) First, calculate the grayscale distribution of the global or local regions of the image. Construct a normalized histogram:
[0061] in, For the Kronecker function, ∈[0,255] represents integer gray levels. Based on this distribution, a cumulative distribution function (CDF) is then constructed:
[0062] b) To address the differences in brightness distribution across different images, an adaptive gamma factor γ is introduced to perform a nonlinear transformation on the CDF:
[0063] in For adjustment coefficients, Represents image entropy; c) Divide the image into several M×M non-overlapping blocks, independently calculate the histogram and CDF transform function of each block, and then perform enhancement:
[0064] in, This is the modulated mapping function. It is adaptively determined by statistics such as the histogram / CDF of the pixel's neighborhood, local contrast, entropy, or gradient gating, and is used for local contrast stretching or brightness remapping. It is the fusion weight of each local block for pixel (x,y); d) To prevent over-enhancement and artifacts near the target edges, an edge gradient-aware control mechanism is introduced:
[0065] For pixel locations where the gradient value is greater than a set threshold τ, reduce the enhancement amplitude:
[0066] Among them, edge adaptive weights Defined as:
[0067] β controls the edge suppression intensity, ensuring edge continuity and structural integrity. The resulting enhanced infrared image is then obtained. .
[0068] Key points: 1) Optical flow-guided infrared image sequence alignment mechanism: This invention designs a four-layer pyramid-shaped optical flow estimation network based on the U-Net structure to calculate the forward and reverse optical flow fields in the image sequence, respectively, for explicit inter-frame registration, which significantly improves the alignment accuracy and robustness of multi-frame fusion.
[0069] 2) Multi-scale temporal mutual attention structure: In view of the structural complexity and thermal field change characteristics of infrared images, this invention divides the feature channel into three sub-channels and constructs attention branches with different temporal downsampling rates (ρ=1, 2, 3) to effectively capture short-term dynamic and long-term structural information and achieve "soft alignment" temporal information modeling.
[0070] 3) Optical flow driven output enhancement module: Based on multi-scale temporal mutual attention output, this invention uses the optical flow fields of the previous and next frames to re-align the output features and fuse the enhancement values of the current frame with those of the previous and next frames to further improve the consistency and stability of the features and make up for the response deviation of the attention mechanism under fast motion conditions.
[0071] 4) Dynamic Nonlinear Gray-Level Stretching Module: A histogram enhancement method combining global and local gray-level distributions was designed. Through gamma modulation function, adaptive image entropy adjustment, local block fusion, and edge gradient protection mechanisms, it achieves significant enhancement of the target hot area while suppressing background noise and enhancement artifacts. 5) End-to-end structural coupling and embedded deployment friendliness: The processing flow from input image to final enhanced image can be jointly optimized end-to-end, and the computational complexity of each module is uniform, supporting modular deployment, which is suitable for embedded implementation of infrared processing chips for aviation platforms.
Claims
1. A method for denoising and enhancing infrared video images, characterized in that, The method includes: Step 1: Perform grayscale normalization and size adjustment on the acquired continuous infrared image sequence to obtain a standardized image sequence. ; Step 2: Construct a pyramid-shaped framework for the continuous infrared image sequence after grayscale normalization and resizing. Process it; Step 3: Based on the pyramid-shaped framework, extract the continuous infrared image sequence. The optical flow characteristics; wherein, the optical flow characteristics include a forward optical flow field. and reverse optical flow field ; Step 4: Construct a hierarchical denoising network for continuous infrared image sequences. Denoising is performed on mid-infrared images to obtain feature sequences; Step 5: Based on the forward optical flow field and reverse optical flow field The feature sequence is enhanced to obtain the enhanced feature sequence; Step 6: Based on the enhanced feature sequence, a dynamic nonlinear grayscale enhancement module based on histogram modeling is introduced to improve the contrast and structural clarity of the infrared image after noise removal, and enhance the visual perception of thermal targets in the image.
2. The method according to claim 1, characterized in that, Step 1 includes: Acquiring continuous infrared image sequences And perform grayscale normalization and size adjustment. This adapts to subsequent network input; where H is the image height and W is the image width.
3. The method according to claim 1, characterized in that, Step 2 includes: a) Input frame and The pieces are spliced along the channel dimension to form a size of Input tensor ; b) In the first layer, for the input tensor By two consecutive times After convolution and GELU activation, preliminary edge and texture information is extracted to obtain a 64-channel feature map. Subsequently, max pooling is used for spatial downsampling, reducing the size to half of the original, i.e., the image height is H / 2 and the image width is W / 2. c) The second layer uses the same double convolutional structure as before to extract higher-level contextual information, and further reduces the spatial resolution through max pooling. The image height becomes H / 4, the image width becomes W / 4, and the feature dimension is increased to 128 channels. d) The third layer continues to apply the double residual structure downsampling, the image height becomes H / 8, the image width becomes W / 8, and a 256-channel feature map is obtained, which has the ability to represent the deformation of objects and the global structure. e) The fourth layer is the deepest layer. The image height becomes H / 16 and the image width becomes W / 16. Double convolution is used to increase the feature dimension to 512 channels while preserving the global semantic features with the lowest spatial resolution. f) The feature map of the fifth layer is first subjected to bilinear upsampling to restore the size to twice that of the previous layer, the image height becomes H / 8, the image width becomes W / 8, and the channel dimension is reduced to 256 through a convolution operation; then it is concatenated with the feature map of the third layer symmetrical to the encoder to fuse high-level semantics and low-level detail information. g) Repeat the same operations for each subsequent level: upsampling, convolution, and concatenation; the corresponding fused layers are the second layer and the first layer, with convolution kernels of all... The output channels are 128 and 64 respectively; h) In the last layer, the high-resolution feature map obtained by the decoder is mapped to a shape of size through an output convolutional layer. The two-dimensional optical flow field has two channels that correspond to the horizontal and vertical motion of each pixel in the image.
4. The method according to claim 1, characterized in that, Step 3 includes: a) will By inputting the above framework, the forward optical flow field is obtained. ; b) Then By inputting the above framework, the reverse optical flow field is obtained. .
5. The method according to claim 1, characterized in that, Step 4 includes: a) Infrared image sequence The input is fed into a convolutional encoder for feature stretching. This encoder consists of multiple two-dimensional convolutional layers and outputs a feature tensor. in, This is the feature encoder. This step yields the temporal feature sequence for feature expansion: in, The total number of feature channels is pre-divided into three sub-channels according to the modeling requirements of different scales; b) Next, this feature tensor The features are split into three sub-channels based on the channel dimension, denoted as follows: Each feature first passes through a The convolutional kernel performs channel compression and dimension alignment, serving as a consistent input for subsequent attention calculations. c) The three features mentioned above are respectively input into three temporal mutual attention modules; each module contains a multi-head self-attention computation structure and introduces different temporal downsampling rate parameters ρ: The first branch uses ρ=1, indicating that the entire frame participates in attention modeling; The second branch uses ρ=2, and performs temporal sampling every 2 frames; The third branch uses ρ=3, sampling every 3 frames, focusing on long-term dependency modeling; d) In a pair of images, each serves as both a reference frame and a supporting frame. and They serve as each other's reference and support frames, in order to As a reference frame, To support frames For reference frame Its features are passed through a set of learnable linear mapping weights. Convert to query vector : For each frame in the support frames Keys are generated using either shared or independent weights. AND value : in, , respectively, are the mapping matrices for keys and values, and d is the internal dimension set in the attention calculation; e) Any frame The information aggregation form in the attention mechanism is as follows: in, This represents the set of supported frames selected based on the downsampling rate ρ. , Attention weights are determined by the reference frame. With supporting frames Calculated; All attention operations are learnable and optimized end-to-end during training; f) Features of each attention output Each input is a set of feedforward neural network (FFN) modules for information augmentation. The structure is typically as follows: First layer: convolution + GELU activation; Second layer: convolution + residual connection, expressed by the formula: g) The enhanced features of the three branches are as follows: They are spliced together along the channel dimension: in, The first step is to perform a dimensional concatenation operation, followed by feature compression and fusion using a 1×1 convolutional kernel to restore the channel dimension to 1, resulting in the final output. 。 6. The method according to claim 1, characterized in that, Step 5 includes: a) For the feature sequence output by the temporal mutual attention module Furthermore, an enhancement mechanism based on optical flow guidance is introduced to improve the consistency and continuity of cross-frame features, especially to suppress detail blurring or target drift caused by factors such as attention mismatch and inter-frame deformation. b) Utilizing the forward optical flow field from step 3 and reverse optical flow field The output features of the preceding and following frames are transformed to the coordinate system of the current frame, respectively, to obtain the forward enhancement features and the backward enhancement features: in, This represents a bilinear interpolation operation based on optical flow, used to map features from different frames to their spatiotemporal locations in a reference frame. These are forward enhancement features and backward enhancement features, respectively; c) Output features of the current frame The features of adjacent frames aligned with it are fused to form a final feature representation with stronger structural consistency: d) Compress the final feature representation through a 1×1 convolution: e) Finally, the enhanced feature sequence is obtained. : 。 7. The method according to claim 1, characterized in that, Step 6 includes: a) First, calculate the grayscale distribution of the global or local regions of the image. Construct a normalized histogram: in, For the Kronecker function, ∈[0,255] represents integer gray levels. Based on this distribution, a cumulative distribution function (CDF) is then constructed: b) To address the differences in brightness distribution across different images, an adaptive gamma factor γ is introduced to perform a nonlinear transformation on the CDF: in, For adjustment coefficients, Represents image entropy; c) Divide the image into several M×M non-overlapping blocks, independently calculate the histogram and CDF transform function of each block, and then perform enhancement: in, The modulated mapping function, It is the fusion weight of each local block for pixel (x,y); d) To prevent over-enhancement and artifacts near the target edges, an edge gradient-aware control mechanism is introduced: For pixel locations where the gradient value is greater than a set threshold τ, reduce the enhancement amplitude: 。 8. The method according to claim 7, characterized in that, Edge adaptive weights Defined as: β controls the edge suppression intensity, ensuring edge continuity and structural integrity, ultimately resulting in an enhanced infrared image. .