High-definition night vision driving data enhancement method based on deep learning

Abstract

The invention discloses a high-definition night vision driving data enhancement method based on deep learning. The method comprises the following steps: S1, constructing a paired image sample set; s2, inputting the paired image sample set into a double-branch depth encoder network; s3, performing information interaction and fusion on the features extracted by the first branch and the second branch to obtain a fusion feature map; s4, performing target sensitivity modeling on the fused feature map, constructing a target sensitivity score map of an image region, and taking image block-level features as a search space; s5, introducing an improved badger algorithm to carry out adaptive optimization on the target sensitivity score map, and outputting an optimized target sensitivity score map; s6, inputting the optimized target sensitivity score chart into a condition image generator to generate a night vision image; and S7, discriminating the night vision image, optimizing generator parameters, and outputting an enhanced night vision image. According to the method, the night vision image is enhanced and generated by combining the multi-modal depth features and the improved badger algorithm.

Description

Technical Field

[0001] The present invention relates to the technical fields of deep learning and image processing, and in particular to a high-definition night vision driving data enhancement method based on deep learning. Background Art

[0002] With the continuous development of intelligent driving, assisted driving, and unmanned vehicle technologies, the visual perception ability of vehicles in low-illumination environments such as at night, in tunnels, and in rain and fog has become one of the core technologies for ensuring driving safety. In this context, night vision imaging and image enhancement technologies have received increasing attention. Existing night vision systems mainly rely on infrared thermal imaging technology or low-light camera systems, but both have limitations. Infrared images can perceive heat source targets in the absence of light, have good penetration and target detection capabilities, but due to the lack of texture details, their image quality is low, the scene structure is blurred, and it is difficult to provide complex semantic information. In contrast, low-light visible light images can provide rich texture and edge details, but they have high noise, low contrast, and poor recognition rate in low-light or no-light environments. How to fuse these two image modalities, give full play to their respective advantages, and achieve high-quality and high-resolution night vision image enhancement is a key issue in the current field of night driving perception.

[0003] Traditional infrared and visible light image fusion methods mostly use means such as pixel-level weighting, gradient domain fusion, and wavelet transform at the image level. Although these methods can achieve basic information superposition in some static scenes, in the actual driving environment, due to problems such as asynchronous sensor positions, different acquisition frequencies, and image noise interference, the fusion effect is extremely unstable, and there is a lack of adaptive attention ability to the target area, which easily leads to the weakening of key targets or the enhancement of background interference. At the same time, traditional algorithms are difficult to adapt to rapidly changing traffic scenes, are not sensitive to moving targets, and cannot meet the requirements of real-time performance and robustness.

[0004] In recent years, with the development of deep learning, especially architectures such as generative adversarial networks (GANs) and Transformers, significant breakthroughs have been achieved in image enhancement and image generation tasks. Existing research has begun to attempt to extract features from infrared and visible light images using convolutional neural networks and perform end-to-end image reconstruction. Some methods introduce conditional generators or attention mechanisms to improve the quality of image generation. However, these methods still have many deficiencies. First, most existing multimodal image fusion networks adopt general backbone structures and do not make structural divisions for the heat source characteristics in infrared images and the texture information in visible light images, resulting in redundant fusion features and blurred target boundaries. Second, most mainstream attention mechanisms are simple weightings in the channel or spatial directions and do not have the ability to search for regions, making it difficult to accurately locate key target regions. Third, existing methods rarely introduce dynamic interpretable mechanisms to simulate the behavior logic in the real world, such as biological perception or environmental response paths, lacking controllability and self-adaptability. Fourth, most deep models are large in volume and require high computing resources, making them unsuitable for deployment on edge computing platforms for real-time inference and difficult to meet the strict requirements of night vision systems for latency and energy consumption.

[0005] In addition, existing research generally ignores the construction of a regional attention mechanism during the image generation process. In complex environments, a single full-image reconstruction method often enhances unnecessary details in non-target regions, which instead weakens the perception effect of target regions (such as pedestrians, vehicles, obstacles, etc.).

[0006] Therefore, how to provide a high-definition night vision driving data enhancement method based on deep learning is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose a high-definition night vision driving data enhancement method based on deep learning. The present invention integrates a multimodal deep feature extraction network and an improved honey badger algorithm with a dynamic search mechanism, and realizes the efficient fusion of infrared and visible light images and the clear enhancement of night vision images by constructing image block-level target sensitivity modeling and guided conditional generation. It has the advantages of strong target focusing, high image structure restoration, flexible model deployment, and good real-time inference, and is applicable to intelligent driving scenarios in low-light environments.

[0008] The high-definition night vision driving data enhancement method based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Use in-vehicle sensors to synchronously collect infrared images and visible light images, complete image alignment through timestamp registration and optical flow estimation, and construct a paired image sample set;

[0010] S2. Input the paired image sample set into the dual-branch deep encoder network. The first branch uses the ResNet-50 network to extract the heat source boundary features of the infrared image, and the second branch uses the Swin Transformer to extract the texture features of the visible light image;

[0011] S3. Perform information interaction and fusion on the heat source boundary features of the infrared image extracted by the first branch and the texture features of the visible light image extracted by the second branch to obtain a fused feature map;

[0012] S4. Perform target sensitivity modeling on the fused feature map, construct a target sensitivity score map for the image region, and use the image patch-level features as the search space;

[0013] S5. Introduce an improved honey badger algorithm to adaptively optimize the target sensitivity score map and output the optimized target sensitivity score map;

[0014] S6. Use the optimized target sensitivity score map as a condition to input into the conditional image generator to generate a night vision image;

[0015] S7. Discriminate the night vision image and optimize the parameters of the conditional image generator, and deploy it to the edge computing device to perform real-time inference, and finally output an enhanced night vision image.

[0016] The present invention proposes a set of high-definition night vision driving data enhancement methods based on deep learning, establishing a complete process from image acquisition, feature extraction, information fusion, sensitivity modeling, adaptive optimization, image generation to edge deployment, effectively solving the problems of poor image quality, difficult target recognition, and insufficient deployment ability of existing night vision images in low-light environments. This method fuses the heat source characteristics of infrared images and the texture features of visible light images, realizes deep fusion through multi-branch coding and cross-attention, introduces target sensitivity score modeling at the image patch level, and combines an improved honey badger algorithm that simulates biological behavior for regional focusing optimization, and finally guides the image generator to generate night vision images with clear structures and real details. This method has strong robustness, adaptability and real-time performance, can be efficiently deployed at the edge, and has important application value in night application scenarios such as autonomous driving.

[0017] Optionally, the S1 specifically includes:

[0018] S11. Synchronously collect the infrared image I ir and the visible light image I vis through in-vehicle sensors deployed on the driving platform. The infrared image obtains thermal radiation information in the wavelength range of 8 micrometers to 15 micrometers, and the visible light image obtains visible light texture data in the range of 400 nanometers to 700 nanometers;

[0019] S12. Add frame-level timestamps to the infrared image and the visible light image respectively and where t represents the sequence number of the acquisition time. Set the time synchronization tolerance ε to perform frame-level registration, so that it satisfies Achieve time-domain consistency;

[0020] S13. For the registered image pair Perform gray normalization processing to unify the pixel value range to [0,1];

[0021] S14. Use the dense optical flow estimation algorithm to perform pixel-level motion modeling on the image pair to generate an optical flow field where u t (x,y) represents the offset of the pixel point (x,y) in the horizontal direction, and v t (x,y) represents the offset of the pixel point (x,y) in the vertical direction;

[0022] S15. Transform the position of the pixel points in the infrared image through the optical flow field to obtain the aligned infrared image The mapping relationship is:

[0023]

[0024] S16. Combine the aligned infrared image and the corresponding visible light image to form a paired image sample And summarize to form a paired image sample set where n represents the number of frames of the acquired images.

[0025] By defining in detail the process of infrared image and visible light image acquisition, alignment and paired sample set construction, the present invention ensures the time synchronization and spatial consistency of the input data, effectively avoiding the image mismatch problem caused by sensor acquisition errors. By adopting the method of timestamp registration combined with optical flow estimation, high-precision alignment of different modality images at the frame level and pixel level is achieved, improving the accuracy and consistency of subsequent feature fusion. Especially the image remapping method driven by optical flow enables the infrared image to be accurately mapped to the visible light image coordinate space, further enhancing the collaborative expression ability of multi-modal data, laying a high-quality data foundation for subsequent deep fusion and enhancement processing, and improving the overall performance and stability of the system.

[0026] Optionally, the ResNet-50 network includes, from the input layer to the output layer in sequence, an initial convolutional layer with a convolutional kernel size of 7×7 and a stride of 2, a max pooling layer, four residual convolutional groups Block1 to Block4. The residual convolutional groups are respectively set with a fixed number of residual modules, and each layer in the residual module adopts a structural configuration with the same number of input and output channels.

[0027] By adopting the ResNet-50 network structure in the first branch to extract the heat source boundary features of the infrared image, the structural perception ability of the model for heat source targets in low-light scenarios is effectively improved. ResNet-50 has the ability of deep feature extraction and the advantage of residual connection, which can capture weak thermal signals and edge contours in the infrared image and enhance the model's response to the target area. At the same time, the unified channel design and standardized configuration of each residual block in the network ensure the stable propagation and fusion of features at different levels, avoiding the occurrence of gradient disappearance or feature collapse problems. This structure strengthens the model's attention ability to significant regions in the infrared modality, significantly improves the accuracy and robustness of target extraction, and provides a distinguishable feature basis for subsequent multi-modal fusion.

[0028] Optionally, the Swin Transformer includes a linear embedding module, window multi-head self-attention modules in multiple stages, and a normalization processing module;

[0029] The linear embedding module divides the visible light image into image patches of size P×P, and performs a linear transformation on each image patch to obtain an initial feature representation, and the dimension of the initial feature representation is:

[0030]

[0031] where H represents the height of the visible light image, W represents the width of the visible light image, and C represents the number of channels;

[0032] The window multi-head self-attention module uses a non-overlapping window division method to divide the feature representation in each stage into several local windows of M×M, and alternately introduces a shifted window mechanism in the odd and even layers, and the shift size is used to enhance the cross-window modeling ability;

[0033] Perform multi-head self-attention calculation within each window, and after normalizing the output of each window through the normalization processing module, perform connection combination.

[0034] By introducing the Swin Transformer structure to perform multi-level modeling on the texture features of the visible light image, the model's ability to extract complex texture and structural information is improved. Swin Transformer performs the self-attention mechanism within local windows and realizes cross-region feature interaction through the shifted window strategy, taking into account both local detail retention and global information aggregation, and overcoming the problem of limited receptive fields of traditional convolutional models. This structure can fully capture weak textures, edge structures and blurred details in low-light images, and enhance the model's ability to analyze and utilize semantic features in visible light images. Window normalization and multi-stage hierarchical processing further optimize the continuity and stability of feature expression, providing a rich and stable source of texture information for the subsequent fusion process.

[0035] Optionally, S3 specifically includes:

[0036] S31. Unify the spatial dimensions and channel dimensions of the heat source boundary features of the infrared image extracted by the first branch and the texture features of the visible light image extracted by the second branch;

[0037] S32. Perform a channel mapping operation on the heat source boundary features of the infrared image to generate a query feature map for feature fusion:

[0038] Q ir = F ir · W Q ;

[0039] where Q ir represents the query feature map, F ir represents the heat source boundary features of the infrared image, and W Q represents the weight matrix in the query direction;

[0040] S33. Perform a channel mapping operation on the texture features of the visible light image to generate a key feature map and a value feature map for feature fusion respectively:

[0041] K vis = F vis · W K , V vis = F vis · W V ;

[0042] where K vis represents the key feature map, V vis represents the value feature map, F vis represents the texture features of the visible light image, W K represents the weight matrix in the key direction, and W V represents the weight matrix in the value direction;

[0043] S34. Calculate a cross-attention weight matrix based on the query feature map and the key feature map, and apply the cross-attention weight matrix to the value feature map to obtain a fused intermediate feature map in a pixel-by-pixel weighted combination manner:

[0044]

[0045] where F fused represents the fused intermediate feature map, Softmax represents normalization processing, and d represents the dimensional constant of the key;

[0046] S35. Perform channel fusion and spatial dimension reconstruction on the fused intermediate feature map to obtain a fused feature map.

[0047] By constructing a unified spatial dimension and channel mapping mechanism to guide the cross-attention fusion of infrared and visible light feature maps, the deep alignment and information sharing of multi-modal features are achieved. The adopted query, key, and value structures clearly distinguish the roles of the guiding channel and the response channel, ensuring the target directivity and texture response accuracy during the feature interaction process. The cross-attention mechanism enables the infrared features to dynamically adjust the attention area according to the visible light information, effectively improving the identification ability of the fused features for key regions. Finally, through channel fusion and dimension reconstruction, a fused feature map with consistent structure and clear semantics is obtained, providing a high-quality representation with structural guiding significance for the sensitivity modeling and generation tasks.

[0048] Optionally, S4 specifically includes:

[0049] S41. Divide the fused feature map into multiple image patches to form an image patch set where B k represents the k-th image patch, and N b represents the total number of image patches;

[0050] S42. Generate a spatial attention response map for each image patch:

[0051]

[0052] where S k (i, j) represents the attention response value at the position of the i-th row and the j-th column in the k-th image patch, σ represents the activation function, C represents the number of channels, ω c represents the weighting coefficient of the c-th channel, B k (c, i, j) represents the pixel value at the position of the c-th channel in the i-th row and the j-th column in the k-th image patch, represents the non-linear mapping function, ψ represents the position mapping function, and P k (i, j) represents the two-dimensional relative spatial position encoding at the position of the i-th row and the j-th column in the k-th image patch;

[0053] S43. Use the attention response map to perform per-pixel weighting processing on the image patch and calculate the initial target sensitivity score of the image patch:

[0054]

[0055] where, represents the initial target sensitivity score of the k-th image patch, γ c represents the fusion weight parameter of the c-th channel, h represents the height of the image patch, and w represents the width of the image patch;

[0056] S44. Normalize the initial target sensitivity scores of all image patches to obtain the final target sensitivity score;

[0057] S45. Reconstruct the final target sensitivity score into a target sensitivity score map according to the positions of the image patches in the original image.

[0058] After dividing the fusion feature map into image patches, the target sensitivity modeling method proposed in the above steps establishes a spatial attention response map, and constructs an explicit saliency scoring mechanism by combining a non-linear mapping function and position encoding. Through the saliency weighting and normalization process at the image patch level, the local response enhancement of the target region is achieved. This mechanism enables the model to have the ability to identify potential target regions, without relying on manual annotation or prior information, and has high adaptability. The generation of the sensitivity score map provides clear regional guidance for subsequent path optimization and conditional generators, effectively improving the fidelity and resolution control ability of the target region in the image generation process, and having good engineering generality and scalability.

[0059] Optionally, the S5 specifically includes:

[0060] S51. Use the generated fusion feature map, image patch set, and target sensitivity score map as the input of the improved honey badger algorithm, and initialize the honey badger individual position set in the image patch space. where M represents the number of honey badger individuals, represents the search path position of the i-th honey badger individual in the target sensitivity score map, and all positions are initialized by weighted sampling with the target sensitivity score as the probability distribution;

[0061] S52. Define the fitness function of the improved honey badger algorithm as:

[0062]

[0063] where, represents the fitness function, represents the search path position of the i-th honey badger individual at the t-th iteration, represents the target sensitivity score of the search path position of the i-th honey badger individual at the t-th iteration, represents the edge gradient variance of the image patch at the search path position of the i-th honey badger individual at the t-th iteration, represents the direction consistency index between the search path position of the i-th honey badger individual at the t-th iteration and the previous moment path, and λ1, λ2, and λ3 represent weight parameters, satisfying λ1 + λ2 + λ3 = 1;

[0064] S53. In each iteration, update the long-range search direction and jump step size regulation of the honey badger individual based on the sound propagation mechanism, and the update expression is:

[0065]

[0066] Among them, represents the jump step size of the i-th honey badger individual at the t-th iteration, β0 represents the initial jump step size, exp represents the exponential function, and γ represents the decay factor. represents the Euclidean distance between the search path position of the current honey badger individual and the global optimal position. The sound propagation mechanism guides the jump exploration by dynamically regulating the jump step size.

[0067] S54. Control the local path refinement operation based on the odor localization mechanism. The path update expression is:

[0068]

[0069] Among them, represents the search path position of the i-th honey badger individual at the (t + 1)-th iteration, sign represents the sign function. represents the global optimal position. represents the perturbation matrix of the target sensitivity score map, which is used to simulate the odor distribution to guide the directional convergence.

[0070] S54. Iterate until the maximum number of iterations T max or the fitness convergence threshold is reached, and extract the global optimal path set.

[0071] S55. According to the global optimal path set perform block-level weighted reconstruction on the target sensitivity score map to generate an optimized target sensitivity score map:

[0072]

[0073] Among them, represents the target sensitivity score of the updated k-th image block, α k represents the target sensitivity score of the k-th image block before update, η represents the enhancement factor, I represents the Boolean indicator function, and X k represents the spatial coordinate position of the k-th image block. If the image block position hits the global optimal path set, 1 is output; otherwise, 0 is output.

[0074] By taking the set of image patches, the sensitivity score map, and the fused feature map as inputs together, an improved honey badger algorithm is introduced to achieve adaptive path optimization. The proposed fitness function fuses the sensitivity score, the edge gradient variance, and the direction consistency metric, and combines the sound propagation and odor localization mechanisms to guide individuals to achieve a dynamic balance between local intensification and global exploration in the image space. The jump step size and the perturbation matrix are continuously adjusted during iteration, so that the final convergence path focuses on regions with significant semantics and clear boundaries. On the basis of keeping the algorithm lightweight, this method simulates the visual path planning behavior of nocturnal organisms, breaks through the static limitation of the traditional attention mechanism, and significantly enhances the focusing ability on the target region and the controllability of image expression.

[0075] Optionally, S6 specifically includes:

[0076] S61. Concatenate the optimized target sensitivity score map and the fused feature map to construct a conditional input containing guiding information;

[0077] S62. Introduce the conditional input into the conditional image generator as a control signal to regulate the attention region response during the image generation process;

[0078] S63. Inside the conditional image generator, use the conditional input for feature modeling and pixel reconstruction stages to guide the allocation of target region information;

[0079] S64. Generate a night vision image at the output end of the conditional image generator, and the night vision image contains enhanced detail regions and boundary features after conditional regulation.

[0080] By introducing the target sensitivity score map as a conditional input into the image generator, an image reconstruction process with explicit region guidance is constructed, changing the way of homogeneous reconstruction of the entire image in the traditional image generation process. The conditional image generator can focus on allocating generation resources according to the target regions indicated in the sensitivity map, making the generated image have the characteristics of clear structure and enhanced boundaries. The guidance mechanism plays a role in the feature modeling and pixel reconstruction stages, effectively suppressing problems such as background interference and image blurring, and improving the saliency of the target region and the degree of detail retention. This structure enhances the interpretability and controllability of the model for the image generation process through a structured guidance path, and improves the expression quality of night vision images in complex environments.

[0081] Optionally, S7 specifically includes:

[0082] S71. Input the night vision image generated by the conditional image generator and real night image samples into the discriminant module, and the discriminant module discriminates the authenticity of the night vision image at different spatial resolutions;

[0083] S72. Construct a combined loss function to guide the optimization of the parameters of the conditional image generator;

[0084] S73. According to the error signal output by the combined loss function, use the backpropagation algorithm to update the parameters in the conditional image generator;

[0085] S74. After completing the training and optimization, export the conditional image generator model and convert it into an inference format suitable for edge hardware, and complete the freezing and quantization processing of the conditional image generator model structure;

[0086] S75. Deploy the processed conditional image generator model to the edge computing device, and initialize the input interface and cache structure;

[0087] S76. In the edge computing device, receive image data and conditional inputs in real time, complete the automatic generation inference process of night vision images, and finally output enhanced night vision images.

[0088] By constructing a complete generation-discrimination optimization closed-loop, high-quality training and efficient inference deployment of the image enhancement model are achieved. The introduction of the discrimination module provides a realistic judgment criterion for the generated images. The combined loss function combines objectives such as structural similarity and edge preservation to constrain the quality of image generation from multiple dimensions. After training, the model undergoes freezing and quantization processing, featuring a streamlined structure and efficient computation, and can be deployed on edge devices for real-time inference. This process realizes end-to-end optimization from model performance to deployment execution, ensuring the speed and quality of enhanced image generation while meeting the low-latency and high-reliability requirements of scenarios such as autonomous driving, and improving the industrial applicability and deployment flexibility of the system.

[0089] Optionally, the combined loss function includes a generative adversarial loss, a structural similarity loss, and an edge preservation loss. The generative adversarial loss is used to minimize the difference between the generated night vision image and the real night image in the discriminator output. The structural similarity loss is used to maintain the consistency of the brightness, contrast, and structural information of the night vision image. The edge preservation loss is calculated based on the difference in the gradient maps of the night vision images and is used to enhance the edge detail expression ability of the night vision images.

[0090] The above steps clarify the composition and functions of the combined loss function adopted in the image generation optimization process, ensuring that the image generation results have both authenticity, structural integrity, and detail expressiveness. Among them, the generative adversarial loss improves the overall visual quality of the image, the structural similarity loss maintains the consistency of the original image structure and semantics, and the edge preservation loss enhances the edge detail expression ability. The three cooperate and optimize to make the night vision image have better boundary clarity, noise control ability, and texture restoration level. The design of this loss function is applicable to various complex environments such as low light, haze, and night, significantly improving the robustness of the model and the quality of image output, and providing reliable enhancement support for high-definition night vision scenarios.

[0091] The beneficial effects of the present invention are as follows:

[0092] First, different from the simple superposition or shallow fusion of traditional infrared and visible light images, the present invention uses a dual-branch structured design. ResNet-50 is used for modeling the heat source boundary of infrared images, and Swin Transformer is used for texture semantic extraction of visible light images, enabling the two modalities to complete information differential expression and structured division of labor in the feature extraction stage, thereby improving the representation ability of the fused image in different scenarios. With the introduction of the cross-attention mechanism, deep feature interaction between the two branches is achieved in the encoding stage, further enhancing the semantic consistency and target recognition ability of the fusion.

[0093] Second, the present invention first introduces an image patch-level target sensitivity modeling mechanism in the multi-modal image fusion task. By constructing a local spatial attention response map and calculating the target saliency score, the system can identify the image patches that may contain key target regions, forming an explicit regional perception structure. Combining with the improved honey badger algorithm proposed by the present invention, it simulates the dynamic perception path behavior of organisms at night in a thermal perception environment, uses the sound propagation mechanism to regulate the remote search direction and step size, and combines the odor localization mechanism to complete local focusing fine-tuning, so that the attention weight of the target region gradually converges and is significantly strengthened in the iteration, realizing an adjustable, trackable, and spatially responsive adaptive attention mechanism. This mechanism is significantly different from the attention structures based only on channel weighting or static weight allocation in the prior art, enhancing the focusing ability on the target region and the adaptive performance of the system.

[0094] In the image generation stage, the present invention, based on the conditional image generator structure, first introduces the target sensitivity score map as explicit conditional information into the generation process. Through the conditional injection mechanism, the image generator is guided to focus on sensitive regions and perform structure reconstruction, making the output image have clear boundaries, real textures, and high semantic restoration. At the same time, a joint optimization mechanism composed of multiple loss functions is introduced, including generative adversarial loss, structural similarity loss, perceptual loss, and edge-preserving loss, etc., systematically improving the image generation quality and solving problems such as blurred image reconstruction, missing target structures, and semantic inconsistencies in existing methods.

[0095] In addition, through a lightweight model structure and module pruning strategy, the present invention deploys the complete enhanced model on edge computing devices with neural network processing units, achieving real-time inference capabilities of more than 20 frames per second on the premise of meeting the accuracy requirements, and having the engineering implementation advantages of low latency, low power consumption, and high responsiveness, which are particularly suitable for scenarios such as autonomous driving and in-vehicle assisted driving systems. Brief Description of the Drawings

[0096] The accompanying drawings are used to provide a further understanding of the present invention and form a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the accompanying drawings:

[0097] Figure 1 is the overall flowchart of the high-definition night vision driving data enhancement method based on deep learning proposed by the present invention;

[0098] Figure 2 is the structural schematic diagram of the dual-branch deep encoder network of the high-definition night vision driving data enhancement method based on deep learning proposed by the present invention;

[0099] Figure 3 is the path search and optimization flowchart of the improved honey badger algorithm for the high-definition night vision driving data enhancement method based on deep learning proposed by the present invention. Detailed implementation manners

[0100] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0101] Referring to Figures 1-3 , the high-definition night vision driving data enhancement method based on deep learning includes the following steps:

[0102] S1. Use in-vehicle sensors to synchronously collect infrared images and visible light images, complete image alignment through timestamp registration and optical flow estimation, and construct a paired image sample set;

[0103] S2. Input the paired image sample set into the dual-branch deep encoder network. The first branch uses the ResNet-50 network to extract the heat source boundary features of the infrared image, and the second branch uses the Swin Transformer to extract the texture features of the visible light image;

[0104] S3. Perform information interaction and fusion on the heat source boundary features of the infrared image extracted by the first branch and the texture features of the visible light image extracted by the second branch to obtain a fused feature map;

[0105] S4. Perform target sensitivity modeling on the fused feature map, construct a target sensitivity score map for the image region, and use the image patch-level features as the search space;

[0106] S5. Introduce an improved honey badger algorithm to adaptively optimize the target sensitivity score map and output the optimized target sensitivity score map;

[0107] S6. Use the optimized target sensitivity score map as a condition to input into the conditional image generator to generate night vision images;

[0108] S7. Discriminate the night vision image, optimize the parameters of the conditional image generator, deploy them to the edge computing device to perform real-time inference, and finally output the enhanced night vision image.

[0109] In this embodiment, the S1 specifically includes:

[0110] S11. Synchronously collect the infrared image I ir and the visible light image I vis by the vehicle-mounted sensors deployed on the driving platform. The infrared image obtains the thermal radiation information with a wavelength range of 8 to 15 micrometers, and the visible light image obtains the visible light texture data of 400 to 700 nanometers.

[0111] S12. Add frame-level timestamps to the infrared image and the visible light image respectively and where t represents the sequence number of the acquisition moment, and set the time synchronization tolerance ε to perform frame-level registration so that is satisfied to achieve time-domain consistency.

[0112] S13. Perform gray normalization processing on the registered image pair to unify the pixel value range to [0, 1].

[0113] S14. Use the dense optical flow estimation algorithm to perform pixel-level motion modeling on the image pair to generate an optical flow field where u t (x, y) represents the offset of the pixel point (x, y) in the horizontal direction, and v t (x, y) represents the offset of the pixel point (x, y) in the vertical direction.

[0114] S15. Transform the position of the pixel points in the infrared image through the optical flow field to obtain the aligned infrared image The mapping relationship is:

[0115]

[0116] S16. Combine the aligned infrared image and the corresponding visible light image to form a paired image sample and summarize them to form a paired image sample set where n represents the number of frames of the acquired images.

[0117] In this embodiment, the ResNet-50 network sequentially includes an initial convolutional layer with a convolutional kernel size of 7×7 and a stride of 2, a max pooling layer, and four residual convolutional groups Block1 to Block4 from the input layer to the output layer. The residual convolutional groups are respectively set with a fixed number of residual modules, and each layer in the residual module adopts a structural configuration with the same number of input and output channels.

[0118] In this embodiment, the Swin Transformer includes a linear embedding module, window multi-head self-attention modules of multiple stages, and a normalization processing module;

[0119] The linear embedding module divides the visible light image into image patches of size P×P, and performs a linear transformation on each image patch to obtain an initial feature representation, and the dimension of the initial feature representation is:

[0120]

[0121] where H represents the height of the visible light image, W represents the width of the visible light image, and C represents the number of channels;

[0122] The window multi-head self-attention module adopts a non-overlapping window division method, divides the feature representation in each stage into several local windows of size M×M, and alternately introduces a shifted window mechanism in the odd and even layers, and the shift size is for enhancing the cross-window modeling ability;

[0123] Perform multi-head self-attention calculation within each window, and after normalizing the output of each window through the normalization processing module, perform connection combination.

[0124] In this embodiment, the S3 specifically includes:

[0125] S31. Unify the spatial dimensions and channel dimensions of the heat source boundary features of the infrared image extracted by the first branch and the texture features of the visible light image extracted by the second branch;

[0126] S32. Perform a channel mapping operation on the heat source boundary features of the infrared image to generate a query feature map for feature fusion:

[0127] Q ir =F ir ·W Q ;

[0128] where Q ir represents the query feature map, F ir represents the heat source boundary features of the infrared image, and W Q represents the weight matrix in the query direction;

[0129] S33. Perform a channel mapping operation on the texture features of the visible light image to generate a key feature map and a value feature map for feature fusion respectively:

[0130] K vis =F vis ·W K ,V vis =F vis·W V ;

[0131] Among them, K vis represents the key feature map, V vis represents the value feature map, F vis represents the texture feature of the visible light image, W K represents the weight matrix of the key direction, W V represents the weight matrix of the value direction;

[0132] S34. Calculate the cross-attention weight matrix based on the query feature map and the key feature map, and apply the cross-attention weight matrix to the value feature map to obtain the fused intermediate feature map in a pixel-by-pixel weighted combination manner:

[0133]

[0134] Among them, F fused represents the fused intermediate feature map, Softmax represents the normalization process, and d represents the dimensional constant of the key;

[0135] S35. Perform channel fusion and spatial dimension reconstruction on the fused intermediate feature map to obtain the fused feature map.

[0136] In this embodiment, the specific content of S4 includes:

[0137] S41. Divide the fused feature map into multiple image patches to form an image patch set Among them, B k represents the k-th image patch, and N b represents the total number of image patches;

[0138] S42. Generate a spatial attention response map for each image patch:

[0139]

[0140] Among them, S k (i, j) represents the attention response value at the position of the i-th row and the j-th column in the k-th image patch, σ represents the activation function, C represents the number of channels, ω c represents the weighting coefficient of the c-th channel, B k (c, i, j) represents the pixel value at the position of the c-th channel in the i-th row and the j-th column in the k-th image patch, represents the non-linear mapping function, ψ represents the position mapping function, and P k (i, j) represents the two-dimensional relative spatial position encoding at the position of the i-th row and the j-th column in the k-th image patch;

[0141] S43. Perform pixel-by-pixel weighting processing on the image patch using the attention response map to calculate the initial target sensitivity score of the image patch:

[0142]

[0143] Among them, represents the initial target sensitivity score of the k-th image patch, and γ c represents the fusion weight parameter of the c-th channel, h represents the height of the image patch, and w represents the width of the image patch;

[0144] S44. Normalize the initial target sensitivity scores of all image patches to obtain the final target sensitivity scores;

[0145] S45. Reconstruct the final target sensitivity scores into a target sensitivity score map according to the positions of the image patches in the original image.

[0146] In this embodiment, the specific steps of S5 include:

[0147] S51. Use the generated fusion feature map, the image patch set, and the target sensitivity score map as the input of the improved honey badger algorithm, and initialize the honey badger individual position set in the image patch space where M represents the number of honey badger individuals, represents the search path position of the i-th honey badger individual in the target sensitivity score map, and all positions are initialized by weighted sampling with the target sensitivity score as the probability distribution;

[0148] S52. Define the fitness function of the improved honey badger algorithm as:

[0149]

[0150] Among them, represents the fitness function, represents the search path position of the i-th honey badger individual at the t-th iteration, represents the target sensitivity score of the search path position of the i-th honey badger individual at the t-th iteration, represents the edge gradient variance of the image patch at the search path position of the i-th honey badger individual at the t-th iteration, represents the direction consistency index between the search path position of the i-th honey badger individual at the t-th iteration and the previous moment path, and λ1, λ2, and λ3 represent weight parameters, satisfying λ1 + λ2 + λ3 = 1;

[0151] S53. In each iteration, update the long-range search direction and jump step size regulation of the honey badger individuals based on the sound propagation mechanism, and the update expression is:

[0152]

[0153] Among them, represents the jump step size of the $i$-th honey badger individual at the $t$-th iteration, $\beta_0$ represents the initial jump step size, exp represents the exponential function, and $\gamma$ represents the decay factor. represents the Euclidean distance between the search path position of the current honey badger individual and the global optimal position, and the sound propagation mechanism guides the jump exploration by dynamically regulating the jump step size.

[0154] S54. Control the local path refinement operation based on the odor localization mechanism, and the path update expression is:

[0155]

[0156] where represents the search path position of the $i$-th honey badger individual at the $(t + 1)$-th iteration, sign represents the sign function, represents the global optimal position, represents the perturbation matrix of the target sensitivity score map, which is used to simulate the odor distribution to guide the directional convergence.

[0157] S54. Iterate until the maximum number of iterations $T$ max or the fitness convergence threshold is reached, and extract the global optimal path set

[0158] S55. According to the global optimal path set perform block-level weighted reconstruction on the target sensitivity score map to generate an optimized target sensitivity score map:

[0159]

[0160] where represents the target sensitivity score of the updated $k$-th image block, $\alpha$ k represents the target sensitivity score of the $k$-th image block before update, $\eta$ represents the enhancement factor, $I$ represents the Boolean indicator function, and $X$ k represents the spatial coordinate position of the $k$-th image block. If the image block position hits the global optimal path set, output 1; otherwise, output 0.

[0161] In this embodiment, the specific steps of S6 include:

[0162] S61. Concatenate the optimized target sensitivity score map and the fusion feature map to construct a conditional input containing guiding information.

[0163] S62. Introduce the conditional input into the conditional image generator as a control signal to adjust the attention area response during the image generation process.

[0164] S63. Inside the conditional image generator, the conditional input is used for the feature modeling and pixel reconstruction stages to guide the allocation of the target region information.

[0165] S64. A night vision image is generated at the output end of the conditional image generator. The night vision image includes detail regions and boundary features enhanced after conditional regulation.

[0166] In this embodiment, the specific steps of S7 are as follows:

[0167] S71. The night vision image generated by the conditional image generator and the real night image samples are input into the discriminant module. The discriminant module discriminates the authenticity of the night vision image at different spatial resolutions.

[0168] S72. A joint loss function is constructed to guide the optimization of the parameters of the conditional image generator.

[0169] S73. According to the error signal output by the joint loss function, the backpropagation algorithm is used to update the parameters in the conditional image generator.

[0170] S74. After the training and optimization are completed, the conditional image generator model is exported and converted into an inference format suitable for edge hardware, and the structure freezing and quantization processing of the conditional image generator model are completed.

[0171] S75. The processed conditional image generator model is deployed to the edge computing device, and the input interface and cache structure are initialized.

[0172] S76. In the edge computing device, image data and conditional input are received in real time to complete the automatic generation inference process of the night vision image, and finally an enhanced night vision image is output.

[0173] In this embodiment, the joint loss function includes a generative adversarial loss, a structural similarity loss, and an edge-preserving loss. The generative adversarial loss is used to minimize the difference between the generated night vision image and the real night image at the discriminator output. The structural similarity loss is used to maintain the consistency of the brightness, contrast, and structural information of the night vision image. The edge-preserving loss is calculated based on the difference in the gradient maps of the night vision image and is used to enhance the edge detail expression ability of the night vision image.

[0174] Example 1:

[0175] In order to verify the feasibility of the present invention in implementation, the present invention is applied to an autonomous driving road test scenario jointly built by a national key intelligent driving research center and a new energy vehicle company. Continuous actual measurements were carried out under low-light conditions at night from December 2024 to January 2025. The test vehicle was a modified pure electric autonomous driving test platform equipped with an infrared imager (wavelength range 8-14μm), a low-light CMOS camera (minimum illumination perception 0.01Lux) and an NVIDIAJetson Orin NX edge computing module. The test time covers multiple time periods from dusk to late night. The lighting environment ranges from urban lighting, low-light highways, and open roads without lighting. The weather covers sunny days, foggy days, and light rain.

[0176] On the traditional image enhancement module, the image of the vehicle is blurred when driving at night, and the edges of obstacles are difficult to identify, especially in haze or backlight reflection conditions. Although the infrared image can show the outline of the heat source, the texture structure is completely lost. Although the visible light image has structure, the noise is serious and the image is dark. As a result, the driving system has problems such as misidentification of the target in front and failure of path planning many times. In order to solve the above-mentioned key perception bottleneck problems, the research team connected the high-definition night vision driving data enhancement method based on deep learning proposed in the present invention to the original vehicle visual processing link, and synchronously collected infrared images and visible light images in the form of paired images at the front end of the driving. After image alignment, dual-channel encoding and fusion processing, the target sensitivity score map is constructed based on the image block, and then the improved honey badger algorithm is introduced to complete the regional optimization, guide the generator to generate night vision images, and finally use the enhanced image as the visual input of the perception system for real-time reasoning.

[0177] During the entire test process, the system continuously collected and processed a total of 26,372 frames of night driving image data, including 16,340 frames of low-light images, 5,226 frames of haze environment images, and 4,806 frames of unlit scene images. By comparing with the traditional image enhancement system, it is found that the present invention has significant advantages in maintaining image structure clarity, boundary accuracy and target recognizability. The system uses multiple evaluation indicators such as SSIM (structural similarity index), PSNR (peak signal-to-noise ratio), edge clarity score, and night target detection recall rate for comprehensive analysis.

[0178] The test results show that on unlit roads, the average SSIM of the images enhanced by the traditional method is only 0.55, the PSNR is 17.3 dB, and the average recall rate of target detection is 58.2%; while after using the method of the present invention, the SSIM is significantly increased to 0.83, the PSNR is increased to 22.9 dB, and the average recall rate of target detection is increased to 84.7%. Especially in extremely hazy weather, the present invention enhances the restoration degree of structural textures such as vehicles and road signs while keeping the heat source boundary unchanged, reducing the detection system delay rate to within 82 ms, supporting the edge platform to operate stably at 20 frames per second, and meeting the real-time requirements of autonomous driving.

[0179] In addition, in terms of optimizing the target sensitivity modeling, through the evaluation and statistics at the image block level, it is found that the average weight of the sensitive image blocks containing heat sources or texture cross regions in the present invention is increased by 42.3%, and the average weight suppression rate of the non-target regions reaches 58.7%, indicating that the improved honey badger algorithm achieves significant focusing on the key regions and effectively avoids the amplification of background noise. The test team also observed that after using the present invention, the false alarm rate of the night driving system decreased by more than 31%, and the system stability was significantly improved.

[0180] Table 1 Performance comparison table of the present invention and traditional methods in night vision image enhancement under different environments

[0181]

[0182]

[0183] From the comparison data in Table 1 above, it can be seen that in the urban lighting night scene, the present invention is superior to the traditional method in all key indicators. Specifically, in this scenario, the structural similarity index (SSIM) of the traditional method is 0.61, the peak signal-to-noise ratio (PSNR) is 18.7 dB, the image enhancement delay is 142 milliseconds, and the target detection recall rate is only 64.3%. Under the same conditions, the present invention has the SSIM increased to 0.85, the PSNR increased to 23.5 dB, the recall rate increased to 87.9%, the image enhancement delay significantly decreased to 78 milliseconds, and the edge sharpness score increased from 0.48 to 0.72, indicating that the present invention not only effectively improves the image quality and detail restoration ability, but also has obvious advantages in the model inference efficiency.

[0184] In the night scene of a road without lights, the illumination of the image source is extremely low, and the image quality of traditional methods further deteriorates. The SSIM and PSNR are 0.55 and 17.3 dB respectively, the recall rate is 58.2%, the edge sharpness is only 0.41, and the latency is as high as 151 milliseconds, severely restricting the practicality of the system in a lightless environment. In contrast, the present invention still maintains stable performance in this scenario, with an SSIM of 0.83, a PSNR of 22.9 dB, the recall rate increased to 84.7%, the image enhancement latency controlled within 79 milliseconds, and the edge sharpness score reaching 0.69, fully demonstrating the effective extraction and reconstruction ability of the present invention for target structures and details in an extremely low illumination environment.

[0185] In a light haze environment, traditional methods are prone to image degradation due to the superposition of blurred infrared image contours and visible light image texture noise. Its SSIM drops to 0.52, PSNR drops to 16.1 dB, the recall rate drops to 51.6%, and the image enhancement latency rises to 163 milliseconds. In contrast, the present invention realizes key optimization of sensitive areas by introducing an improved honey badger attention mechanism, with the SSIM increased to 0.81, PSNR increased to 22.4 dB, the recall rate increased to 81.3%, the latency reduced to 83 milliseconds, and the edge sharpness score improved from 0.38 to 0.66, significantly enhancing the image recognizability and system response speed in a complex aerosol environment.

[0186] In the most challenging medium and heavy haze environment, traditional methods perform the worst, with an SSIM of only 0.49, a PSNR of only 15.5 dB, the recall rate dropping to 46.2%, severe loss of edge details, and the enhancement latency increasing to 170 milliseconds. In such scenarios, the target contours are blurred and the background interference is strong, which is a technical bottleneck in night vision enhancement systems. However, the present invention still exhibits strong robustness, with an SSIM reaching 0.78, a PSNR of 21.7 dB, a target recall rate as high as 79.5%, the image enhancement latency controlled within 86 milliseconds, and the edge sharpness score maintained at 0.62, indicating that the sensitive area modeling and adaptive image reconstruction mechanism of the present invention have good adaptability to extreme scenarios.

[0187] Based on the above analysis, the present invention has achieved higher image quality indicators, faster system response speed, and more accurate target recognition ability than traditional methods in various typical night vision low-light environments, comprehensively verifying its practical application value in the night vision enhancement task of autonomous driving. Especially under extreme conditions of no light or severe haze, the target focusing mechanism and regional optimization strategy of the present invention can effectively improve the structural expressiveness and system stability of images.

[0188] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A method for enhancing high-definition night vision driving data based on deep learning, characterized in that, It includes the following steps: S1. Use in-vehicle sensors to synchronously collect infrared images and visible light images, complete image alignment through timestamp registration and optical flow estimation, and construct a paired image sample set; S2. Input the paired image sample set into a dual-branch deep encoder network. The first branch uses a ResNet-50 network to extract the heat source boundary features of the infrared image, and the second branch uses a Swin Transformer to extract the texture features of the visible light image; S3. Perform information interaction and fusion on the heat source boundary features of the infrared image extracted by the first branch and the texture features of the visible light image extracted by the second branch to obtain a fused feature map; S4. Perform target sensitivity modeling on the fused feature map, construct a target sensitivity score map for the image region, and use the image patch-level features as the search space; S5. Introduce an improved honey badger algorithm to adaptively optimize the target sensitivity score map and output the optimized target sensitivity score map; S6. Use the optimized target sensitivity score map as a condition to input into a conditional image generator to generate a night vision image; S7. Discriminate the night vision image and optimize the parameters of the conditional image generator, and deploy it to an edge computing device to perform real-time inference, and finally output an enhanced night vision image.

2. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, characterized in that, The specific content of S1 includes: S11. Synchronously collect the infrared image I and the visible light image I through the vehicle-mounted sensors deployed on the driving platform. The infrared image acquires thermal radiation information with a wavelength range of 8 to 15 micrometers, and the visible light image acquires visible light texture data of 400 to 700 nanometers. ir and the visible light image I vis , the infrared image acquires thermal radiation information with a wavelength range of 8 to 15 micrometers, and the visible light image acquires visible light texture data of 400 to 700 nanometers; S12. Add frame-level timestamps to the infrared image and the visible light image respectively and where t represents the sequence number of the acquisition moment, and a time synchronization tolerance ε is set to perform frame-level registration so that achieve time-domain consistency; S13. For the registered image pair perform gray normalization to unify the pixel value range to [0, 1]; S14. Use the dense optical flow estimation algorithm to perform pixel-level motion modeling on the image pair and generate an optical flow field where u t (x, y) represents the offset of the pixel point (x, y) in the horizontal direction, and v t (x, y) represents the offset of the pixel point (x, y) in the vertical direction; S15. Perform position transformation on the pixel points in the infrared image through the optical flow field to obtain the aligned infrared image The mapping relationship is as follows: S16. Combine the aligned infrared image and the corresponding visible light image to form a paired image sample And summarize to form a paired image sample set Where n represents the number of frames of the acquired images.

3. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, wherein, The ResNet-50 network successively includes an initial convolutional layer with a convolutional kernel size of 7×7 and a stride of 2, a max pooling layer, and four residual convolutional groups Block1 to Block4 from the input layer to the output layer. The residual convolutional groups are respectively set with a fixed number of residual modules, and each layer in the residual module is configured with a structure with the same number of input and output channels.

4. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, wherein The Swin Transformer includes a linear embedding module, window multi-head self-attention modules in multiple stages, and a normalization processing module; The linear embedding module divides the visible light image into image patches with a size of P×P, and performs a linear transformation on each image patch to obtain an initial feature representation; The window multi-head self-attention module adopts a non-overlapping window partitioning method, divides the feature representation in each stage into several local windows of M×M, and alternately introduces a shifted window mechanism in the odd and even layers, with a shift size of to enhance the cross-window modeling ability; Perform multi-head self-attention calculations within each window, and after normalizing the output of each window through the normalization processing module, perform connection combination.

5. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, characterized in that, The specific content of S3 includes: S31. Unify the spatial dimensions and channel dimensions of the heat source boundary features of the infrared image extracted by the first branch and the texture features of the visible light image extracted by the second branch; S32. Perform a channel mapping operation on the heat source boundary features of the infrared image to generate a query feature map for feature fusion; S33. Perform a channel mapping operation on the texture features of the visible light image to respectively generate a key feature map and a value feature map for feature fusion; S34. Calculate a cross-attention weight matrix based on the query feature map and the key feature map, and apply the cross-attention weight matrix to the value feature map to obtain a fused intermediate feature map in a pixel-by-pixel weighted combination manner; S35. Perform channel fusion and spatial dimension reconstruction on the fused intermediate feature map to obtain a fused feature map.

6. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, wherein, The specific content of S4 includes: S41. Divide the fused feature map into multiple image patches to form an image patch set where B k represents the k-th image patch, and N b represents the total number of image patches; S42. Generate a spatial attention response map for each image patch: Among them, S k (i, j) represents the attention response value at the position of the i-th row and the j-th column in the k-th image patch, σ represents the activation function, C represents the number of channels, ω c represents the weighting coefficient of the c-th channel, B k (c, i, j) represents the pixel value of the c-th channel at the position of the i-th row and the j-th column in the k-th image patch, represents the non-linear mapping function, ψ represents the position mapping function, P k (i, j) represents the two-dimensional relative spatial position encoding of the k-th image patch at the position of the i-th row and the j-th column; S43. Perform pixel - by - pixel weighting on the image patches using the attention response map, and calculate the initial target sensitivity score of the image patches: Among them, represents the initial target sensitivity score of the k-th image patch, and γ c represents the fusion weight parameter of the c-th channel, h represents the height of the image patch, and w represents the width of the image patch; S44. Normalize the initial target sensitivity scores of all image patches to obtain the final target sensitivity score; S45. Reconstruct the final target sensitivity score into a target sensitivity score map according to the positions of the image patches in the original image.

7. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, characterized in that The specific steps of S5 are as follows: S51. Use the generated fused feature map, image patch set, and target sensitivity score map as the input of the improved honey badger algorithm, and initialize the honey badger individual position set in the image patch space. where M represents the number of honey badger individuals. represents the search path position of the i-th honey badger individual in the target sensitivity score map, and all positions are initialized by weighted sampling with the target sensitivity score as the probability distribution. S52. Define the fitness function of the improved honey badger algorithm as: Among them, represents the fitness function, represents the search path position of the i-th honey badger individual at the t-th iteration, represents the target sensitivity score of the search path position of the i-th honey badger individual at the t-th iteration, represents the edge gradient variance of the image block at the search path position of the i-th honey badger individual at the t-th iteration, represents the direction consistency index between the search path position of the i-th honey badger individual at the t-th iteration and the previous moment path, and λ1, λ2, and λ3 represent weight parameters, satisfying λ1 + λ2 + λ3 = 1; S53. In each iteration, update the long - range search direction and jump step - size regulation of the honey badger individuals based on the sound propagation mechanism. The update expression is: Among them, represents the jump step size of the i-th honey badger individual at the t-th iteration. β0 represents the initial jump step size, exp represents the exponential function, and γ represents the decay factor. represents the Euclidean distance between the search path position of the current honey badger individual and the global optimal position. The sound propagation mechanism guides the jump exploration by dynamically adjusting the jump step size. S54. Control the local path refinement operation based on the odor localization mechanism. The path update expression is: Among them, represents the search path position of the i-th honey badger individual at the (t + 1)-th iteration, sign represents the sign function, represents the global optimal position, represents the perturbation matrix of the target sensitivity score map, which is used to simulate the odor distribution to guide directional convergence; S54. Iterate until the maximum number of iterations T max or the fitness convergence threshold is reached, and extract the set of globally optimal paths S55. According to the global optimal path set perform block-level weighted reconstruction on the target sensitivity score map to generate an optimized target sensitivity score map: Among them, represents the target sensitivity score of the k-th image block after update, and α k represents the target sensitivity score of the k-th image block before update, η represents the enhancement factor, I represents the Boolean indicator function, and X k represents the spatial coordinate position of the k-th image block, and outputs 1 if the image block position hits the global optimal path set, otherwise outputs 0.

8. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, wherein The specific steps of S6 are as follows: S61. Concatenate the optimized target sensitivity score map and the fusion feature map to construct a conditional input containing guiding information; S62. Introduce the conditional input into the conditional image generator as a control signal to regulate the attention region response during the image generation process; S63. Inside the conditional image generator, use the conditional input for feature modeling and pixel reconstruction stages to guide the allocation of target region information; S64. Generate a night - vision image at the output end of the conditional image generator. The night - vision image contains enhanced detail regions and boundary features after conditional regulation.

9. The method for enhancing high-definition night vision driving data based on deep learning according to claim 1, wherein The specific steps of S7 are as follows: S71. Input the night - vision image generated by the conditional image generator and the real - night - image samples into the discriminant module. The discriminant module discriminates the authenticity of the night - vision image at different spatial resolutions; S72. Construct a joint loss function to guide the parameter optimization of the conditional image generator; S73. According to the error signal output by the joint loss function, use the backpropagation algorithm to update the parameters in the conditional image generator; S74. After completing training and optimization, export the conditional image generator model and convert it into an inference format suitable for edge hardware, and complete the freezing and quantization processing of the conditional image generator model structure; S75. Deploy the processed conditional image generator model to the edge computing device, and initialize the input interface and cache structure; S76. In the edge computing device, receive image data and conditional input in real - time, complete the automatic generation inference process of the night - vision image, and finally output the enhanced night - vision image.

10. The method for enhancing high-definition night vision driving data based on deep learning according to claim 9, wherein, The joint loss function includes a generative adversarial loss, a structural similarity loss, and an edge - preserving loss. The generative adversarial loss is used to minimize the difference between the generated night - vision image and the real - night - image at the discriminator output. The structural similarity loss is used to maintain the consistency of the brightness, contrast, and structural information of the night - vision image. The edge - preserving loss is calculated based on the difference in the gradient maps of the night - vision image and is used to enhance the edge detail expression ability of the night - vision image.

Images