Image dynamic range adaptive correction method in new energy vehicle visual perception

Abstract

The application discloses a new energy vehicle visual perception image dynamic range adaptive correction method, relates to the technical field of image processing and intelligent visual perception, and comprises the following steps: acquiring original image data from a vehicle-mounted camera of a new energy vehicle, extracting a preliminary feature map through a convolutional neural network, determining a potential key region boundary in the image, and obtaining a feature description containing brightness distribution, dynamic range distribution and object contour; the new energy vehicle visual perception image dynamic range adaptive correction method can improve the adaptability of an output image to a new energy vehicle visual perception system decision module, thereby improving the accuracy, stability and reliability of visual perception results in a complex road environment.

Description

Technical Field

[0001] This invention relates to the fields of image processing and intelligent visual perception technology, specifically to an image dynamic range adaptive correction method for visual perception in new energy vehicles. Background Technology

[0002] With the rapid development of new energy vehicle technology, in-vehicle vision perception systems have been widely applied in scenarios such as assisted driving, environmental recognition, path planning, and safety warning. As a crucial information source for vision perception systems, the image quality acquired by in-vehicle cameras directly affects the accuracy of target recognition, scene understanding, and subsequent decision-making modules. Especially in complex road environments, the brightness distribution, dynamic range, and target contours of different regions in the image vary significantly. If the image is not effectively corrected and optimized, it can easily lead to inaccurate visual perception results, thereby affecting the safety and reliability of the entire vehicle system.

[0003] In existing technologies, image enhancement and correction for vehicles typically employ methods such as overall brightness adjustment, global dynamic range compression, unified filtering enhancement, or fixed parameter mapping to improve image display and enhance subsequent recognition capabilities. Under normal lighting conditions, these methods can improve image quality to some extent. However, in the actual operation of new energy vehicles, the environment in which the vehicle operates is significantly dynamic and complex. Scenarios such as tunnel entrances and exits, nighttime roads, backlighting, strong light reflection, rain and fog, and areas with alternating light and dark areas can all result in the simultaneous presence of low-brightness areas, high-brightness areas, and target areas with significant differences in contour details within the same image frame. In such cases, a uniform processing approach for the entire image often fails to meet the imaging needs of different areas. Furthermore, existing technologies generally lack the ability to effectively distinguish the importance of regions within an image, failing to differentiate processing based on regions that truly have decision-making value in visual perception tasks. For example, during the operation of new energy vehicles, areas such as vehicles ahead, pedestrians, traffic signs, and lane boundaries are usually more critical for decision-making, while the sky, distant background, or some unrelated road areas are relatively less important. If existing methods fail to allocate image processing resources reasonably based on the importance, brightness, and contour features of regions, they are prone to problems such as insufficient enhancement of key regions, loss of local details, blurred edge information, while secondary regions are over-processed. This not only reduces image processing efficiency but also affects the visual perception system's recognition and judgment of key targets. Furthermore, existing technologies still have shortcomings in balancing local region optimization with overall image coordination. Even if some methods can correct local dark or bright areas, they often lack comprehensive consideration of the transition relationship between the corrected region and the background region, easily leading to obvious local processing traces, discontinuous edges, and poor fusion effects, thus affecting the naturalness and perceptual consistency of the entire image. Especially when the local enhancement results need to be used for subsequent decision analysis, if key features in the image are incomplete, edge strength is insufficient, or region fusion is unreasonable, it may lead to a deviation in the visual perception system's understanding of the scene. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive correction method for the dynamic range of images in the visual perception of new energy vehicles, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an image dynamic range adaptive correction method for visual perception in new energy vehicles, comprising:

[0006] S1. Obtain raw image data from the on-board camera of the new energy vehicle, extract preliminary feature maps through convolutional neural network, determine the boundaries of potential key regions in the image, and obtain feature descriptions including brightness distribution, dynamic range distribution and object contours.

[0007] S2. Determine the importance of the region based on the feature description. If the brightness distribution is lower than the first preset threshold, it is marked as a low-brightness region. If the brightness distribution is higher than the second preset threshold, it is marked as a high-brightness region. The weight allocation of the low-brightness region or the high-brightness region is enhanced through the attention mechanism to obtain a weighted region priority list.

[0008] S3. Obtain a subset of pixels in high-priority regions using a weighted region priority list. Process these pixel subsets through adaptive filtering and brightness mapping to determine the optimized detail enhancement image blocks and local dynamic range correction image blocks.

[0009] S4. Extract edge intensity information from the optimized detail enhancement image patch and the local dynamic range correction image patch, determine whether the edge intensity exceeds the preset threshold, and if it does, fuse it into the overall image frame to obtain the intermediate image with preliminary correction.

[0010] S5. Calculate the global brightness distribution and dynamic range distribution through the intermediate image, adjust the correction depth of the secondary region according to the brightness distribution and dynamic range distribution, and determine the simplified background pixel set.

[0011] Preferably, S1 includes:

[0012] A preliminary feature map is obtained by extracting spatial information from the original image matrix using a convolutional neural network.

[0013] The boundaries of key regions are determined based on the gradient change rate of the preliminary feature map.

[0014] A brightness distribution histogram is obtained by statistically analyzing the gray-level frequency of the key region boundary. If the extreme value difference of the brightness distribution histogram is greater than a preset threshold, a dynamic range distribution matrix is ​​obtained.

[0015] By fusing the brightness distribution histogram, the dynamic range distribution matrix, and the object contour lines extracted based on the key region boundaries, a feature description containing brightness distribution, dynamic range distribution, and object contour is obtained.

[0016] Preferably, S2 includes:

[0017] The initial brightness distribution map is obtained based on the feature description;

[0018] If the initial brightness distribution map is lower than the first preset threshold, it is marked as a low-brightness area; if it is higher than the second preset threshold, it is marked as a high-brightness area.

[0019] An attention mechanism is used to perform feature mapping on the low-brightness region and the high-brightness region to obtain a weight allocation vector;

[0020] The weighted calculation is performed based on the weight allocation vector, and the weighted region priority list is obtained by sorting. The importance of the region is then determined based on the weighted region priority list.

[0021] Preferably, S3 includes:

[0022] Construct a weighted region priority list based on the initial image blocks, and extract a subset of pixels from the high-priority regions in the weighted region priority list;

[0023] An adaptive filter is used to process the pixel subset to obtain a smooth pixel subset;

[0024] A brightness reconstruction pixel subset is obtained by processing the smooth pixel subset using a nonlinear mapping function constructed based on the smooth pixel subset;

[0025] High-frequency edge information is extracted from the brightness reconstructed pixel subset to determine the optimized detail-enhanced image block, and the brightness reconstructed pixel subset is corrected to determine the local dynamic range corrected image block.

[0026] Preferably, S4 includes:

[0027] Edge extraction is performed on detail-enhanced image patches and local dynamic range-corrected image patches to obtain local edge intensity information;

[0028] Determine whether the local edge intensity information exceeds a preset edge intensity threshold;

[0029] If the local edge intensity information exceeds the edge intensity threshold, then the corresponding pixel is determined to be a high-intensity edge pixel.

[0030] Obtain the global mapping position of the high-intensity edge pixels in the overall image frame;

[0031] Based on the global mapping position, the high-intensity edge pixels are frame-fused to obtain a preliminary corrected intermediate image.

[0032] Preferably, S5 includes:

[0033] An intermediate image is acquired, the intermediate image is processed to obtain a global brightness distribution, and a dynamic range distribution is obtained based on the global brightness distribution.

[0034] If the contrast of a candidate region in the intermediate image is lower than the threshold set by the dynamic range distribution, the candidate region is determined to be a secondary region.

[0035] The depth adjustment factor is calculated based on the global brightness distribution and the dynamic range distribution, and the initial depth of the secondary region is updated using the depth adjustment factor to determine the correction depth.

[0036] The pixels in the secondary region are processed according to the correction depth to determine the simplified background pixel set.

[0037] Preferably, the process also includes S6: obtaining the fusion parameters of the simplified background pixel set and the high-priority region, and integrating the two through pixel-level stitching to obtain the final dynamic range adaptive correction image, specifically including:

[0038] Extract the background pixel set and high-priority target region from the initial acquired image, and remove redundant data from the background pixel set to obtain a simplified background pixel set;

[0039] The fusion ratio is calculated based on the difference in edge gradient between the simplified background pixel set and the high-priority target region.

[0040] Preferably, S6 further includes:

[0041] The simplified background pixel set and the high-priority target region are spliced ​​together using the fusion ratio value to obtain a spliced ​​image matrix;

[0042] The stitched image matrix is ​​mapped using a gamma correction algorithm to obtain the final dynamic range adaptive correction image.

[0043] Preferably, it also includes S7: extracting decision-related features from the final dynamic range adaptively corrected image, judging the feature completeness, and if complete, outputting it to the decision module of the new energy vehicle visual perception system to obtain an adaptively enhanced visual perception result, specifically including:

[0044] The texture distribution is extracted from the dynamic range adaptively corrected image to obtain the initial decision feature set;

[0045] The feature completeness value is calculated based on the initial decision feature set, and the complete decision feature or the repaired decision feature is determined based on the feature completeness value.

[0046] Preferably, S7 further includes:

[0047] The complete decision features or the repair decision features are input into the decision module to generate perception control commands;

[0048] The state mapping is performed according to the perception control command to obtain an adaptively enhanced visual perception result.

[0049] As can be seen from the above technical solution, the present invention has the following beneficial effects:

[0050] This adaptive dynamic range correction method for image dynamic range in new energy vehicle visual perception extracts features from the raw image data acquired by the vehicle-mounted camera. It then identifies potential key regions in the image by combining brightness distribution, dynamic range distribution, and object contours. Based on the importance of each region, low-brightness regions, high-brightness regions, and other key regions are weighted and prioritized. This avoids the problems of insufficient enhancement of key regions, redundant processing of secondary regions, and unreasonable allocation of computational resources caused by applying a uniform correction method to the entire image in existing technologies. Furthermore, by performing adaptive filtering, brightness mapping, edge intensity extraction, and fusion processing on high-priority regions, the method improves the detail clarity, edge continuity, and local dynamic range performance of key target regions. Simultaneously, by adjusting the correction depth of secondary regions and fusing them with high-priority regions, the method balances the naturalness, consistency, and processing efficiency of the overall image. In addition, by extracting decision-related features from the final corrected image and performing integrity judgment, the method improves the adaptability of the output image to the decision-making module of the new energy vehicle visual perception system, thereby enhancing the accuracy, stability, and reliability of visual perception results in complex road environments. Attached Figure Description

[0051] Figure 1 This is a flowchart of the image dynamic range adaptive correction method of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] like Figure 1 As shown, the present invention provides a technical solution: an image dynamic range adaptive correction method for visual perception in new energy vehicles, comprising:

[0054] S1. Obtain raw image data from the on-board camera of the new energy vehicle, extract preliminary feature maps through convolutional neural network, determine the boundaries of potential key regions in the image, and obtain feature descriptions including brightness distribution, dynamic range distribution and object contours.

[0055] S2. Determine the importance of the region based on the feature description. If the brightness distribution is lower than the first preset threshold, it is marked as a low-brightness region. If the brightness distribution is higher than the second preset threshold, it is marked as a high-brightness region. The weight allocation of the low-brightness region or the high-brightness region is enhanced through the attention mechanism to obtain a weighted region priority list.

[0056] S3. Obtain a subset of pixels in high-priority regions using a weighted region priority list. Process these pixel subsets through adaptive filtering and brightness mapping to determine the optimized detail enhancement image blocks and local dynamic range correction image blocks.

[0057] S4. Extract edge intensity information from the optimized detail enhancement image patch and the local dynamic range correction image patch, determine whether the edge intensity exceeds the preset threshold, and if it does, fuse it into the overall image frame to obtain the intermediate image with preliminary correction.

[0058] S5. Calculate the global brightness distribution and dynamic range distribution through the intermediate image, adjust the correction depth of the secondary region according to the brightness distribution and dynamic range distribution, and determine the simplified background pixel set.

[0059] S6. Obtain the fusion parameters of the simplified background pixel set and the high-priority region, and integrate the two through pixel-level stitching operation to obtain the final dynamic range adaptive correction image.

[0060] S7. Extract decision-related features from the final dynamic range adaptively corrected image, determine the completeness of the features, and if complete, output them to the decision module of the new energy vehicle visual perception system to obtain the adaptively enhanced visual perception result.

[0061] In the above implementation, brightness statistical analysis and gradient analysis are performed on the feature map to generate brightness distribution and dynamic range distribution, which are then combined with target detection contour information to form a composite feature description. Based on this feature description, a dual-threshold determination mechanism is introduced to divide the image region into low-brightness regions, high-brightness regions, and normal regions. Furthermore, an attention mechanism (e.g., a structure combining channel attention and spatial attention) is used to enhance the weight of regions with extreme lighting conditions, enabling the model to prioritize these regions in subsequent processing.

[0062] In the above scheme, adaptive filtering (such as bilateral filtering or guided filtering) is used for high-priority regions to preserve edge structure. Simultaneously, a nonlinear brightness mapping function (such as logarithmic mapping or gamma correction) is used to expand the local dynamic range, thereby generating detail-enhanced image patches and dynamic range-corrected image patches. Furthermore, edge intensity information is extracted using edge detection operators (such as the Sobel or Canny operators), and effective enhancement regions are selected based on a threshold, then fused into the global image to form an intermediate image.

[0063] In the above implementation, by performing global statistical analysis on the intermediate image, the overall brightness distribution function and dynamic range index are obtained, and the processing intensity of the background region is dynamically adjusted accordingly. Only key regions are processed in detail, while secondary regions are simplified, thereby reducing computational complexity. Finally, high-priority regions are integrated with the background region through pixel-level fusion strategies (such as weighted fusion or multi-resolution fusion) to output the final image, and features for autonomous driving decision-making (such as target position, boundary information, and semantic labels) are extracted to complete the visual perception closed loop.

[0064] S1 includes extracting spatial information of the original image matrix through a convolutional neural network to obtain a preliminary feature map; determining the boundary of a key region based on the gradient change rate of the preliminary feature map; obtaining a brightness distribution histogram by statistically analyzing the gray-level frequency of the boundary of the key region; obtaining a dynamic range distribution matrix if the extreme value difference of the brightness distribution histogram is greater than a preset threshold; and fusing the brightness distribution histogram, the dynamic range distribution matrix, and the object contour lines extracted based on the boundary of the key region to obtain a feature description containing brightness distribution, dynamic range distribution, and object contour.

[0065] In one implementation, the original image matrix originates from a single-frame road scene image captured by a camera mounted on a new energy vehicle during driving. This original image matrix first enters a convolutional neural network for spatial information extraction. Specifically, the process involves: first, performing size unification and pixel value normalization on the original image matrix to maintain consistent data scale under different imaging conditions, thereby reducing fluctuations in subsequent feature extraction stages. Size unification is determined based on the input specifications of the vehicle-mounted visual perception system, typically determined by the camera output resolution, processing chip cache capacity, and the input size of the subsequent recognition module, ensuring consistency in data size across the pre- and post-processing chains. Pixel value normalization is determined based on the original grayscale range of the image, aiming to compress numerical differences between images with different exposures, making the convolutional layer's response to brightness changes more stable. After preprocessing, the image enters the pre-convolutional layer of the convolutional neural network. The pre-convolutional layer locally scans adjacent pixel regions in the image, extracting brightness transitions, texture direction, edge density, and local structure distribution information layer by layer, thereby forming a preliminary feature map. The number of convolutional layers is determined based on scene complexity, the number of target types, and onboard computing resources. When the scene is highly variable and there are many target categories, a deeper number of layers is chosen; when the scene is relatively simple and chip resources are limited, a shallower number of layers is chosen. The kernel size is determined based on the target scale and the requirement to preserve edge details. When the target is small and the requirement for high contour detail is high, a smaller kernel size is chosen; when the target scale spans a large range, a smaller kernel is used in the front stage to extract details, and the subsequent stage extracts higher-level semantic features by expanding the receptive region. After layer-by-layer convolution processing, the resulting preliminary feature map is no longer a simple collection of pixels, but rather a feature representation that includes spatial relationships, local grayscale variations, and edge response intensity, providing a foundation for subsequent key region boundary extraction.

[0066] After obtaining the preliminary feature map, the boundaries of key regions are determined based on the gradient change rate of the preliminary feature map. Specifically, the process involves performing directional change analysis on the neighborhood of each pixel in the preliminary feature map, examining the grayscale changes in the horizontal and vertical directions, and then combining the intensity of changes in different directions to obtain the gradient change degree at each location. The gradient change rate is essentially a quantification of the rate of grayscale change in adjacent regions; the more drastic the change, the closer the location is to an object boundary, a strong reflective boundary, a shadow boundary, or the outer contour of a road sign. To avoid misjudgments caused by local noise, the preliminary feature map is smoothed and noise-suppressed before calculating the gradient change rate, preventing isolated noise points and random jitter from being identified as valid boundaries. The smoothing intensity is determined based on the image noise level, camera sensor performance, and shooting environment. In nighttime or rain / fog environments with high noise levels, the smoothing intensity is appropriately increased; in sunny, close-up scenes with rich details, the smoothing intensity is appropriately decreased to avoid weakening effective edges. Subsequently, the gradient change rate at each location is compared with the boundary determination threshold. When the gradient change rate is higher than the threshold, the corresponding location is considered a boundary candidate point. The boundary determination threshold is determined based on the gradient statistics of the real boundary regions in the training samples. First, the change level at clear target boundaries is statistically analyzed, and then the change level in non-boundary flat regions is statistically analyzed. The value with higher distinguishability between the two types of regions is taken as the initial threshold, and it is corrected in combination with the recognition stability under scenarios such as daytime, nighttime, backlight, and tunnel entrances and exits, so that the boundary candidate points do not introduce too much background noise, nor do they miss key contours. After obtaining the boundary candidate points, connectivity screening, isolated point removal, and broken edge completion are performed to retain candidate points that are spatially continuous, morphologically reasonable, and consistent with the target structure of the scene, forming the key region boundary. The integrity of the key region boundary is judged by a comprehensive evaluation of the boundary continuity length, boundary transition rationality, and neighborhood grayscale consistency. Candidate boundaries with excessively short continuity length, abnormal transitions, or chaotic surrounding grayscale distribution are excluded, so that the extraction results are closer to the real road target and illumination boundary structure.

[0067] After determining the boundaries of the key regions, a brightness distribution histogram is obtained by statistically analyzing the gray-level frequency of the key region boundaries. The specific implementation process is as follows: First, a predetermined width is extended inwards and outwards along the boundaries of the key regions to form a boundary analysis band. Then, the gray levels of each pixel within this boundary analysis band are statistically analyzed and categorized according to gray-level intervals to obtain the pixel quantity distribution results corresponding to each gray-level interval. The width of the boundary analysis band is determined based on the relationship between the target edge thickness, camera resolution, and scene depth. When the resolution is high and the target edge is relatively clear, the analysis band width is smaller to highlight the brightness changes near the boundary. When there are many distant targets or the image is blurry, the analysis band width is appropriately increased to cover a more complete transition area. The granularity of gray-level interval division is determined based on the brightness resolution requirements and real-time processing load. A finer granularity results in a more accurate brightness distribution depiction, while a coarser granularity reduces the statistical computation burden. The brightness distribution histogram obtained through the above statistics reflects the distribution of pixels from dark to bright near the boundaries of the key regions. If the brightness distribution is concentrated in the low grayscale range, it indicates that the area is generally dark; if the brightness distribution is concentrated in the high grayscale range, it indicates that the area is generally bright; if the brightness distribution spans multiple grayscale ranges, it indicates that the area has obvious light and dark alternation characteristics. Subsequently, the maximum and minimum grayscale values ​​in the brightness distribution histogram are extracted, and the difference between them is calculated to characterize the overall brightness span of the area. When this extreme value difference is greater than a preset threshold, the dynamic range distribution matrix is ​​obtained. The preset threshold is determined based on the statistical results of the brightness span of key areas in a large number of vehicle driving images. First, the grayscale span of normal lighting areas, slightly backlit areas, and strong contrast areas are statistically analyzed separately. Then, a value sufficient to distinguish between ordinary brightness fluctuations and significant dynamic range imbalances is selected as the judgment boundary, and further adjusted based on camera exposure capabilities, image bit depth, and the detail preservation requirements of the visual perception task. When the threshold is set too low, ordinary brightness fluctuations will be judged as abnormal dynamic range, resulting in an excessively large processing range. When the threshold is set too high, the real high-contrast areas will be difficult to enter the key processing range in time. Therefore, the threshold needs to be repeatedly adjusted according to the calibration image to keep the false positive rate and false negative rate in balance.

[0068] After determining that the extreme value difference of the brightness distribution histogram exceeds the preset threshold, a dynamic range distribution matrix is ​​further generated. The specific implementation process is as follows: the image area covered by the key region boundary is divided into multiple local sub-regions. Then, the highest gray level, lowest gray level, gray level uniformity, and intensity of brightness variation between adjacent pixels are statistically analyzed within each local sub-region to determine the dynamic range state of each local sub-region. The size of the local sub-region is determined based on the complexity of the target structure, edge density, and on-board real-time requirements. When the target contour is dense and local brightness variations are frequent, the sub-region size is smaller to improve depiction accuracy; when the scene is relatively flat or processing latency requirements are high, the sub-region size is larger to reduce computational load. After statistical analysis of each local sub-region, its dynamic range state is rearranged according to its original spatial position to form a dynamic range distribution matrix. This dynamic range distribution matrix is ​​essentially a spatial representation of the brightness span distribution within the key region, reflecting both which parts are severely overly dark or overly bright, and which parts belong to the region with the most intense brightness-dark boundary. To ensure the stability of the dynamic range distribution matrix, after local statistics are completed, consistency correction is performed on the abrupt changes between adjacent sub-regions to prevent isolated outliers from disrupting the overall distribution continuity. The intensity of the consistency correction is determined based on the smoothness of the brightness transition between adjacent sub-regions and the spatial continuity of the road scene. The correction intensity is higher for flat roads and large areas of sky, and lower for densely populated areas with target boundaries, in order to preserve necessary local differences.

[0069] After obtaining the brightness distribution histogram and dynamic range distribution matrix, object contour lines are extracted based on the boundaries of the key regions. The specific implementation process is as follows: The edge extension direction is traced along the boundaries of the key regions, and boundary points with significant continuous responses and positional changes conforming to the object contour rules are concatenated to form candidate contour lines. Subsequently, the candidate contour lines undergo breakpoint connection, burr removal, and shape regularization processing to make the contour lines as continuous and smooth as possible, and to match the edges of the real target. The breakpoint connection distance is determined based on image resolution, target motion blur level, and edge integrity requirements; the connection distance is smaller when the resolution is high, and appropriately increased when blurring is significant. The burr removal scale is set based on the boundary noise size and effective detail size; the removal scale is larger when there are many noise particles, and smaller when the target contour details are rich, to avoid accidentally deleting real edges. After the contour lines are extracted, the closure degree, directional continuity, and local smoothness of the contour lines are reviewed, and abnormal lines that clearly do not conform to the contour rules of road targets, vehicle targets, pedestrian targets, or traffic facilities are excluded. This process preserves the object outlines, which not only reflect the spatial shape of key objects in the scene but also the correspondence between the lighting boundary and the entity boundary.

[0070] After obtaining the brightness distribution histogram, dynamic range distribution matrix, and object contour lines, these three are fused to obtain a feature description containing brightness distribution, dynamic range distribution, and object contour. The specific implementation process is as follows: First, scale unification and position alignment are performed on the three types of information to ensure that statistical information, spatial distribution information, and structural contour information correspond within the same image coordinate system. The scale unification method is determined based on the mapping relationship between the output size of the preliminary feature map and the original image size, ensuring that the information extracted at different processing stages can correspond one-to-one in spatial location. After position alignment, fusion weights are assigned according to the importance of each type of information to the visual perception task. The weight corresponding to the brightness distribution histogram is used to reflect the influence of the overall brightness distribution of the region on dynamic range judgment; the weight corresponding to the dynamic range distribution matrix is ​​used to reflect the influence of local brightness span on abnormal region localization; and the weight corresponding to the object contour lines is used to reflect the influence of maintaining entity boundaries on the stability of subsequent perception results. The weights for each element are determined based on the combined performance of target recognition accuracy, boundary integrity preservation, and brightness / darkness anomaly detection rate in the sample scene. First, the contributions of each of the three types of information are examined individually on the calibration samples. Then, they are combined and corrected according to their contribution levels, ensuring that the fused feature description accurately reflects lighting issues while maintaining the integrity of target structural information. After weight configuration, brightness statistics are mapped to the corresponding spatial regions, dynamic range distribution results are overlaid on the corresponding local locations, and object contour lines are superimposed onto the same feature framework, thus forming a feature description for subsequent dynamic range correction processing. This feature description is no longer limited to a single brightness judgment but simultaneously includes regional brightness / darkness status, local dynamic range strength, and object boundary structure information, providing more sufficient data for subsequent region importance judgments.

[0071] S2 includes obtaining an initial brightness distribution map based on feature description; if the initial brightness distribution map is lower than a first preset threshold, it is marked as a low-brightness region, and if it is higher than a second preset threshold, it is marked as a high-brightness region; an attention mechanism is used to perform feature mapping on the low-brightness region and the high-brightness region to obtain a weight allocation vector; a weighted calculation is performed based on the weight allocation vector, and the regions are sorted to obtain a weighted region priority list; the importance of the regions is determined based on the weighted region priority list.

[0072] In one possible implementation, an initial brightness distribution map is obtained based on a feature description matrix; if the initial brightness distribution map is lower than a first preset threshold, it is marked as a low-brightness region, and if it is higher than a second preset threshold, it is marked as a high-brightness region; an attention mechanism is used to perform feature mapping on the low-brightness region and the high-brightness region to obtain a weight allocation vector; a weighted calculation is performed based on the weight allocation vector, and the regions are sorted to obtain a weighted region priority list; the importance of the regions is determined based on the weighted region priority list.

[0073] In the above implementation, the feature description matrix, as the composite feature result output from the previous processing stage, already contains brightness distribution information, dynamic range distribution information, and object contour information. First, brightness information extraction processing is performed on this feature description matrix. Specifically, the brightness channel data corresponding to each region in the feature description matrix is ​​read, and the brightness values ​​are remapped to a two-dimensional structure consistent with the original image according to their spatial location, thereby generating an initial brightness distribution map. This initial brightness distribution map reflects the distribution of brightness at various locations in the entire image, while maintaining consistency with the boundaries and contour structures of key regions. The brightness values ​​are determined by the encoding information of the gray-level frequency statistics in the feature description matrix, and their value range is determined based on the original gray-level range of the image and the normalization processing method, thus ensuring a uniform scale for brightness expression in different scenes.

[0074] After obtaining the initial brightness distribution map, the brightness value of each pixel or preset region block in the image is compared one by one. When the overall brightness value of a certain region is lower than the first preset threshold, the region is marked as a low-brightness region; when the overall brightness value of a certain region is higher than the second preset threshold, the region is marked as a high-brightness region. The first preset threshold is determined by setting an effective identifiable lower limit of brightness based on statistics from a large number of low-light scene images, and is further modified by combining the camera's lowest resolvable grayscale level, the range of ambient light changes, and the visual perception system's requirements for identifying details in dark areas. This ensures that the threshold can cover real low-brightness areas while avoiding misjudging normal shadow areas as low-brightness areas. The second preset threshold is determined by setting a grayscale statistical result of overexposed areas in high-light or strong light scenes, and is further modified by combining the camera's saturation response range, the upper limit of the image sensor's dynamic range, and the requirements for retaining details in the target area. This ensures that the threshold can effectively identify high-brightness overflow areas while avoiding misjudging normally bright areas that still have details. To improve the stability of region segmentation, a neighborhood consistency check is performed on local regions during threshold determination. If a single pixel meets the condition but its entire neighborhood does not, the pixel is corrected to avoid isolated noise points affecting the region segmentation results.

[0075] After marking the low-brightness and high-brightness regions, an attention mechanism is applied to these regions. Specifically, the process involves first extracting the corresponding sub-region features of the low-brightness and high-brightness regions from the feature description matrix, then remapping these sub-regions to adjust their representation intensity in the overall feature space. The attention mechanism analyzes the contribution of different regions to the visual perception task, enhancing potentially hidden target information in low-brightness regions and potentially overexposed but still important edge information in high-brightness regions. During feature mapping, the brightness deviation of each region is first calculated, i.e., the difference between the region's brightness and the overall average brightness. Simultaneously, the intensity of brightness variations within the region is evaluated based on dynamic range distribution information, and the structural importance of the region is assessed by combining the density of object contour lines. Based on this multi-dimensional information, a corresponding weight value is generated for each region, and the weights are arranged spatially to form a weight allocation vector. The magnitude of the weight allocation vector reflects the priority of the region in subsequent processing, and its determination is based on a comprehensive assessment of the region's brightness anomaly, structural information density, and dynamic range span. The more obvious the brightness anomaly, the richer the contour structure, and the more drastic the dynamic range change, the higher the weight value of the region.

[0076] After obtaining the weight allocation vector, a weighted calculation is performed on each region. Specifically, the weight values ​​in the weight allocation vector are combined with the feature response intensity of the corresponding region, amplifying or suppressing the original feature response according to the weight. For regions with higher weights, their feature responses are enhanced after weighting, thus occupying a more important position in subsequent processing; for regions with lower weights, their feature responses are relatively compressed, reducing their impact on the overall dynamic range correction result. After the weighted calculation is completed, all regions are sorted according to their weighted comprehensive response values, resulting in a weighted region priority list. The sorting process is based on the weighted response intensity of each region from high to low; a higher value indicates a higher priority for that region in image dynamic range correction. The sorting process also incorporates regional spatial continuity correction, merging spatially adjacent regions with similar weights to avoid an overly fragmented priority list.

[0077] Finally, the importance of each region is determined based on a weighted list of region priorities. The specific implementation process is as follows: a priority hierarchy standard is set, dividing the sorting results into high-priority, medium-priority, and low-priority regions. High-priority regions correspond to those requiring priority dynamic range correction and detail enhancement; medium-priority regions correspond to those requiring moderate adjustment; and low-priority regions correspond to those requiring only basic brightness consistency. The priority hierarchy standard is determined based on system real-time requirements, processing resource allocation strategies, and the degree of influence of different regions on visual decision-making. This ensures that critical target regions are processed first, while the processing intensity of background regions is appropriately reduced, thereby reducing the overall computational burden while maintaining processing effectiveness.

[0078] S3 includes constructing a weighted region priority list based on the initial image blocks, extracting a subset of pixels from high-priority regions in the weighted region priority list; processing the pixel subset using adaptive filtering to obtain a smoothed pixel subset; processing the smoothed pixel subset using a nonlinear mapping function constructed based on the smoothed pixel subset to obtain a brightness reconstructed pixel subset; extracting high-frequency edge information from the brightness reconstructed pixel subset to determine optimized detail-enhanced image blocks, and correcting the brightness reconstructed pixel subset to determine local dynamic range corrected image blocks.

[0079] In the above implementation, the initial image segmentation process is as follows: the input image is divided into multiple continuous and non-overlapping regions according to spatial location, with each region containing a fixed number of pixels. The size of the region is determined based on image resolution, target scale, and real-time processing requirements. When the image resolution is high and the target size is small, the region size is smaller to improve local processing accuracy; when real-time processing requirements are high, the region size is appropriately increased to reduce the number of regions. After image segmentation, each region is matched with the weighted region priority list obtained in the previous steps, and high-priority regions are selected based on the priority ranking. The selection ratio of high-priority regions is set according to the system processing capability and the importance of the visual task, usually determined by statistical results of the coverage areas of key targets in the training samples, so that the selected regions can cover the main target locations and areas with significant lighting abnormalities. Subsequently, pixels in these high-priority regions are extracted to form pixel subsets. The pixel subset extraction process reads pixel values ​​point by point according to the spatial range of the region, while maintaining the original spatial structure relationship, so that the pixel positions can still be accurately located in subsequent processing.

[0080] After obtaining the pixel subset, adaptive filtering is performed on it. Specifically, for each pixel in the subset, its neighborhood is selected as a reference set. Weighted smoothing is then applied to the current pixel based on the grayscale differences and spatial distances between these neighborhood pixels. The neighborhood range is determined according to the image noise intensity and the need to preserve edge details; when noise is strong, the neighborhood range is appropriately expanded to improve the smoothing effect; when detail requirements are high, the neighborhood range is appropriately reduced to avoid edge blurring. During the filtering process, pixels with small grayscale differences and close spatial distances to the current pixel are assigned higher weights, while pixels with large grayscale differences or large distances are assigned lower weights, thus suppressing noise while preserving edge structure. The specific weight allocation values ​​are set based on the pixel grayscale difference statistics and spatial distance distribution to ensure consistency in smoothing across different brightness regions. After adaptive filtering, a smoothed pixel subset is obtained, which suppresses random noise while retaining the main structural information.

[0081] After obtaining the smooth pixel subset, a nonlinear mapping function is constructed based on the smooth pixel subset, and this function is used to process the smooth pixel subset to obtain the brightness reconstruction pixel subset. The specific implementation process is as follows: First, the brightness distribution of each pixel in the smooth pixel subset is statistically analyzed, including the minimum brightness value, the maximum brightness value, and the distribution density of intermediate brightness. Based on these statistical results, a brightness mapping relationship is constructed so that the original brightness values ​​can be redistributed within the new brightness range after mapping. The construction of the nonlinear mapping function is determined based on the brightness compression and expansion requirements. For low-brightness areas, the brightness increase is increased to make dark details more obvious; for high-brightness areas, the brightness increase rate is reduced to suppress overexposure; for intermediate brightness areas, a smooth transition is maintained to make the overall visual effect natural and continuous. The specific shape of the mapping function is set according to the brightness distribution characteristics. When the brightness distribution is biased towards low values, the increase of the mapping function in the low-brightness segment is increased; when the brightness distribution is biased towards high values, the compression degree of the mapping function in the high-brightness segment is enhanced. The function adjustment range is determined based on the brightness span of the corresponding region in the dynamic range distribution matrix; the larger the span, the more obvious the mapping adjustment intensity. Through the above processing, the pixel values ​​in the smooth pixel subset are redistributed to obtain the brightness reconstruction pixel subset.

[0082] After obtaining the brightness reconstruction pixel subset, high-frequency edge information is extracted from it to determine the optimized detail-enhanced image patch. Specifically, the process involves performing local change analysis on the brightness reconstruction pixel subset to identify locations where brightness changes are significant between pixels. These locations typically correspond to object boundaries, texture details, and structural inflection points. By analyzing the continuity of brightness differences between adjacent pixels, regions with significant changes and spatial coherence are extracted as a set of high-frequency edge information. The edge detection threshold is determined based on the statistical results of brightness changes in real edge regions in the training samples, ensuring that the extracted edges cover the main structural information without introducing excessive noise edges. Subsequently, the pixel regions corresponding to these high-frequency edge information are separated from the brightness reconstruction pixel subset to form the optimized detail-enhanced image patch. This image patch, while maintaining the original spatial structure, highlights key details and boundary features in the image.

[0083] Simultaneously, a subset of pixels reconstructed for brightness is subjected to overall brightness correction to determine local dynamic range corrected image blocks. Specifically, based on the pixel distribution after brightness reconstruction, the brightness within a local area is balanced to achieve a more uniform distribution and more reasonable coverage. During adjustment, the focus is on reducing the saturation of overly bright areas while improving the visibility of overly dark areas, maintaining a continuous brightness transition between adjacent areas. The correction intensity is determined based on the region's position in the dynamic range distribution matrix; regions with more drastic dynamic range changes receive a larger correction amplitude, while regions with gentler changes receive a smaller correction amplitude to avoid overprocessing. Through this process, a local dynamic range corrected image block is obtained, exhibiting a more balanced brightness distribution and providing a foundation for subsequent overall image fusion.

[0084] S4 includes edge extraction of detail-enhanced image blocks and local dynamic range-corrected image blocks to obtain local edge intensity information; determining whether the local edge intensity information exceeds a preset edge intensity threshold; if the local edge intensity information exceeds the edge intensity threshold, determining the corresponding pixel as a high-intensity edge pixel; obtaining the global mapping position of the high-intensity edge pixel in the overall image frame; and performing frame fusion on the high-intensity edge pixel according to the global mapping position to obtain a preliminary corrected intermediate image.

[0085] In the above implementation, both the detail enhancement image block and the local dynamic range correction image block originate from the preprocessing stage, and their spatial positions correspond to those of the original image. First, edge extraction processing is performed on both types of image blocks. Specifically, the process involves scanning the brightness changes at each position within the image block pixel by pixel, calculating the brightness difference between each pixel and its neighboring pixels, and combining this with changes in different directions to obtain the edge response intensity at that pixel position. The neighborhood range is determined based on the image resolution and edge detail requirements. When the resolution is high or the edge detail requirements are high, the neighborhood range is smaller to improve edge localization accuracy; when the image noise is strong, the neighborhood range is appropriately increased to enhance the stability of edge extraction. Through the above processing, local edge intensity information corresponding to the detail enhancement image block and the local dynamic range correction image block is obtained, reflecting the edge salience at each pixel position.

[0086] After obtaining local edge intensity information, a threshold determination process is performed on the edge intensity of each pixel. Specifically, the edge intensity value of each pixel is compared with a preset edge intensity threshold. When the edge intensity value is greater than the threshold, the pixel is considered to belong to a significant edge region. The edge intensity threshold is determined based on statistical results of the intensity distribution of real edge regions in a large number of labeled images. First, the edge intensity range at clear target boundaries is statistically analyzed, then the edge intensity ranges of flat and noisy regions are statistically analyzed. Values ​​that can effectively distinguish between edges and non-edges are selected from these two distributions as the initial threshold. This threshold is then corrected using stability tests under different lighting conditions to ensure stable discrimination capability in low-brightness, high-brightness, and complex lighting scenarios. If the threshold is set too low, noise points are easily misidentified as edges; if the threshold is set too high, real edges may be missed. Therefore, repeated calibration using multiple scene samples is used to achieve a balance.

[0087] When the edge intensity of a pixel exceeds the defined edge intensity threshold, that pixel is identified as a high-intensity edge pixel. To improve edge continuity, connectivity analysis is performed on adjacent high-intensity edge pixels. Spatially continuous and oriented pixel sets are merged into edge regions, while isolated points and short, broken edges are removed to ensure the integrity of the edge structure. The range of connectivity determination is set based on the edge density and image structure complexity. In complex scenes, the connectivity range is appropriately widened, while in simple scenes, it is appropriately tightened to avoid erroneous connections.

[0088] After identifying high-intensity edge pixels, their global mapping positions within the overall image framework are obtained. Specifically, the spatial coordinate information recorded during the image segmentation stage is used to map the local coordinates of each image block back to their corresponding positions in the original image. The mapping relationship is determined by the starting coordinates and region size during image segmentation. The relative position of each pixel within an image block, plus the starting position of that block in the original image, yields the pixel's global position in the overall image. To ensure mapping accuracy, the boundary coordinates and size information of each block are recorded during segmentation and mapped accordingly, ensuring that the local processing results can accurately revert to the overall image space.

[0089] After obtaining the global mapping location, frame fusion processing is performed on high-intensity edge pixels. Specifically, within the overall image frame, the pixel values ​​at corresponding locations are updated according to an edge enhancement strategy, making high-intensity edge regions stand out in the overall image. During fusion, edge information from detail-enhanced image blocks and brightness information from local dynamic range correction image blocks are comprehensively processed to maintain brightness distribution continuity while ensuring edge sharpness. The fusion weights are determined based on edge intensity and their ranking in region priority; regions with higher edge intensity and higher priority have a greater impact on the fusion process. For cases where adjacent regions overlap or have boundaries, smooth transition processing is used to avoid abrupt brightness changes or edge breaks. After fusion, a pre-corrected intermediate image is obtained, which significantly enhances the performance of key edge regions while maintaining overall structural consistency.

[0090] S5 includes acquiring an intermediate image, processing the intermediate image to obtain a global brightness distribution, and obtaining a dynamic range distribution based on the global brightness distribution; if the contrast of a candidate region in the intermediate image is lower than a threshold set by the dynamic range distribution, then the candidate region is determined to be a secondary region; calculating a depth adjustment factor based on the global brightness distribution and the dynamic range distribution, updating the initial depth of the secondary region through the depth adjustment factor to determine a correction depth; and processing the pixels in the secondary region based on the correction depth to determine a simplified background pixel set.

[0091] In the above implementation, the intermediate image is the result of high-intensity edge pixel fusion. This intermediate image has already undergone enhancement of key edge regions and adjustment of local dynamic range, thus its overall structural information is relatively complete, making it suitable as input for global statistical analysis. When processing the intermediate image to obtain the global brightness distribution, all pixels are first traversed and statistically analyzed according to their spatial location. The brightness value of each pixel is then assigned to its corresponding brightness level according to a preset grayscale interval. The number of pixels in each brightness level is then counted to form a brightness distribution result covering the entire image. The division of grayscale intervals is determined based on the image brightness resolution, image bit depth, and onboard processing resources. Overly coarse division weakens the distinguishing ability between brightness levels, while overly fine division increases the statistical burden. Therefore, the number of intervals is usually selected to balance resolution and processing efficiency, taking into account the camera's output accuracy and the response requirements of the sensing module. After completing the brightness statistics for the entire image, the spatial distribution density of each brightness level is further combined to distinguish the proportion of the main road area, sky area, shadow area, and reflective area in the global brightness, thereby obtaining the global brightness distribution. When obtaining the dynamic range distribution based on the global brightness distribution, the span between the highest and lowest brightness levels of the entire image is first determined. Then, the distribution density of the intermediate brightness regions is used to judge whether the transition between light and dark is smooth. If both high and low brightness levels occupy a large proportion and the intermediate levels are sparsely distributed, it indicates that the overall brightness span of the image is large, and the dynamic range distribution is wide. If the brightness is mainly concentrated in the intermediate levels, it indicates that the overall dynamic range of the image is relatively concentrated. To ensure that the dynamic range distribution is not limited to a single global value, multiple candidate regions are divided along the image space. The brightness span and brightness dispersion of each candidate region are statistically analyzed, and the local statistical results are then mapped back to the entire image to form a dynamic range distribution result that takes into account both the overall state and spatial differences.

[0092] In the above implementation, the generation of candidate regions is based on the spatial block division results of the intermediate image, and is modified by combining the road target distribution density, residual information of the priority of the preceding region, and brightness continuity. The size of the candidate region is determined according to the image resolution, target scale, and processing latency requirements. When the resolution is high and there are many small targets at a distance, the size of the candidate region is smaller to improve the accuracy of local judgment; when the scene is flat and the real-time requirements are high, the size of the candidate region is larger to reduce the number of subsequent calculations. When performing contrast statistics on each candidate region, the brightness difference distribution between adjacent pixels in the region is first calculated, and then the difference between the local highest brightness and the local lowest brightness is calculated. At the same time, the diffusion degree of brightness levels within the region is examined to obtain the contrast characterization result of the candidate region. If the contrast of the candidate region is lower than the threshold set by the dynamic range distribution, the candidate region is determined to be a secondary region. The process of determining the threshold is as follows: first, the dynamic range distribution of the entire image is used to determine whether the current scene belongs to a scene with a large brightness-dark span, a scene with a moderate brightness-dark span, or a scene with a small brightness-dark span, and then the contrast judgment benchmark under the corresponding scene is called respectively. For scenes with a large contrast range, the difference between key targets and the background is often more pronounced. Therefore, the threshold for classifying secondary areas is appropriately increased to include areas with only weak contrast. For scenes with a small contrast range, the threshold is appropriately lowered to avoid prematurely classifying ordinary areas with some detail value as secondary areas. The final value of this threshold is determined based on the contrast statistics of the background area, the road edge extension area, and the non-core target area in the calibration sample. By comparing the contrast differences between the real key area and the background area under different weather conditions, time periods, and exposure conditions, a numerical range that can effectively distinguish between primary and secondary areas is selected. This value is then adjusted based on the possibility of misclassification, thus ensuring that the threshold maintains a stable distinguishing ability across multiple scenes.

[0093] In the above implementation, when a candidate region is determined to be a secondary region, a depth adjustment factor is calculated based on the global brightness distribution and dynamic range distribution. The initial depth of the secondary region is updated using the depth adjustment factor, thereby determining the correction depth. Here, the initial depth characterizes the processing level and intensity of the secondary region in subsequent correction, rather than representing physical distance in a spatial ranging sense. The initial depth is set based on the candidate region's position in the overall image, its brightness state, contrast level, and its adjacency relationship with the primary region. Candidate regions located at the image edge, with moderate brightness, low contrast, and far from the primary target region have a lower initial depth setting; candidate regions located near the road extension direction, connected to the primary region, or carrying background structural information have a higher initial depth setting. The calculation process of the depth adjustment factor is as follows: first, the global brightness distribution is used to determine whether the current image is generally dark, bright, or evenly distributed; then, the dynamic range distribution is used to determine whether the local brightness-darkness span is drastic; finally, combined with the brightness level position of the candidate region in the entire image, it is determined whether the region should undergo enhancement, suppression, or a gradual adjustment. If the overall global brightness is too dark and the candidate area brightness is at a low level, the depth adjustment factor is adjusted upwards to maintain basic discernibility in the secondary area. If the overall global brightness is too bright and the candidate area is at a high level, the depth adjustment factor is adjusted downwards to prevent the background highlights from becoming excessively prominent. If the global brightness distribution is relatively balanced and the candidate area only shows localized lack of contrast, the depth adjustment factor is adjusted only slightly to maintain a natural transition in the background area. The value of the depth adjustment factor is determined by the degree of global brightness shift, the strength of the dynamic range, and the degree of contrast loss in the candidate area. The more significant the global shift, the wider the dynamic range, and the weaker the contrast of the candidate area, the larger the depth adjustment magnitude; conversely, the smaller the magnitude. After updating the initial depth with the depth adjustment factor, the corrected depth is obtained. The corrected depth reflects the correction level that the secondary area should be assigned during the background processing stage. A higher level indicates that the area still needs to retain some brightness compensation and structural transition, while a lower level indicates that the area only needs to undergo simplification processing.

[0094] In the above implementation, pixels in secondary regions are processed according to the correction depth to determine a simplified background pixel set. Specifically, all pixels belonging to the same secondary region are grouped and controlled according to the processing level corresponding to the correction depth. For secondary regions with low correction depths, brightness smoothing, noise suppression, and weak texture compression are mainly performed to ensure visual consistency with the surrounding background while reducing unnecessary detail enhancement. For secondary regions with medium correction depths, in addition to brightness smoothing, certain boundary transition information is maintained to avoid abrupt changes at the boundary with the primary region. For secondary regions with high correction depths, the extension of road textures, background outline trends, and large-scale brightness variations are preserved on the basis of simplification processing to maintain overall scene coherence. During pixel processing, the brightness adjustment range is first determined based on the correction depth, then the remaining contrast within the region is used to determine whether to retain local transition details, and finally, the brightness difference with adjacent regions is used to determine whether to perform boundary softening processing. After the above processing, background region pixels that have undergone simplified correction and are suitable for subsequent overall stitching are selected to form a simplified background pixel set. This background pixel set preserves the basic brightness structure and spatial continuity of the background area while reducing the proportion of irrelevant details, providing a stable foundation for subsequent fusion with high-priority areas.

[0095] S6 includes extracting the background pixel set and high-priority target region from the initial acquired image, removing redundant data from the background pixel set to obtain a simplified background pixel set; calculating a fusion ratio value based on the edge gradient difference between the simplified background pixel set and the high-priority target region; performing a stitching operation on the simplified background pixel set and the high-priority target region using the fusion ratio value to obtain a stitched image matrix; and applying a gamma correction algorithm to the stitched image matrix to obtain the final dynamic range adaptive correction image.

[0096] In the above implementation, the initial acquired image is the raw image data obtained by the vehicle-mounted camera of the new energy vehicle in the current driving scenario. The high-priority target region is derived from the results of the prior region priority judgment and local dynamic range correction processing, mainly including road targets ahead, traffic signs, vehicle outlines, pedestrian outlines, and areas with significant illumination changes and structural value that have a significant impact on driving decisions. The extraction of the background pixel set is specifically completed through spatial region separation. That is, based on the boundaries of the high-priority regions obtained in the prior process, the coverage area of ​​the target region is determined in the initial acquired image, and then the remaining pixels outside the coverage area are collected into the background pixel set. To avoid the transition zone near the target boundary being incorrectly classified into the background region, a boundary transition zone of a predetermined width is retained outside the target region, and a neighborhood consistency check is performed on the pixels within the boundary transition zone. The width of the boundary transition zone is determined based on the image resolution, the clarity of the target edge, and the results of the prior edge extraction. When the resolution is high and the outline boundary is clear, the width of the boundary transition zone is smaller; when there is slight blurring or reflective interference at the target edge, the width of the boundary transition zone is appropriately increased to avoid weakening the target outline. Through the above processing, the initial acquired image can be divided into two parts: a high-priority target region and a background pixel set.

[0097] In the above implementation, redundant data is removed from the background pixel set to obtain a simplified background pixel set. The specific process is as follows: First, the background pixel set is traversed, and the brightness uniformity, texture density, local contrast, and edge density of each background sub-region are statistically analyzed. When a region simultaneously satisfies the following characteristics—gradual brightness change, high texture repetition, low local contrast, and weak edge response—it is determined that the region contains a lot of redundant information. Subsequently, a simplification and retention process is performed on the pixels within these redundant regions. This involves retaining representative pixels that reflect the basic brightness levels and spatial transition trends of the region, while reducing repetitive textures and low-value details, thus forming a simplified background pixel set. The redundancy judgment criteria are determined based on the statistical results of the information contribution of the background region in the sample image. If a certain type of background detail contributes little to subsequent target recognition and scene judgment, its redundancy judgment strength is appropriately increased; if a certain type of background region, although not belonging to the core target range, has auxiliary value for scene continuity or road direction perception, more structural information is appropriately retained. Through this processing, the background region still retains its basic spatial levels and brightness continuity, while reducing the proportion of invalid details, thus reducing the computational burden for subsequent stitching processing.

[0098] In the above implementation, the fusion ratio is calculated based on the edge gradient difference between the simplified background pixel set and the high-priority target region. Specifically, the gradient distribution information of the simplified background pixel set and the high-priority target region near the stitching boundary is extracted. The gradient distribution information is obtained by statistically analyzing the intensity of pixel brightness changes along different directions within the boundary's vicinity, characterizing the degree of structural change and brightness abruptness at the boundary transition. The determination of the boundary's vicinity range is based on the target edge thickness, background transition width, and stitching naturalness requirements; values ​​are smaller when the target edge is clear and larger when the background brightness change is gradual. After obtaining the gradient distribution information, the mean gradient intensity, gradient variation dispersion, and gradient continuation trend at corresponding boundary positions are calculated for both the background and target sides. The results on both sides are then compared to obtain the edge gradient difference. A larger edge gradient difference indicates a more significant difference in structural intensity or brightness change rate between the background and target regions, requiring stronger transition control during subsequent stitching to avoid boundary breaks; a smaller edge gradient difference indicates that the two regions are visually similar, allowing for a more direct fusion method during stitching. The fusion ratio is determined based on the edge gradient difference. When the edge gradient difference is large, the participation ratio of transition side pixels is increased to create a smoother transition between the background and the target. When the edge gradient difference is small, the proportion of original structure preservation in the target area is increased to better preserve the edge strength and detail advantages of key areas. The range of the fusion ratio value is determined based on the evaluation results of natural stitching effect and edge continuity in the sample images, and is adjusted in combination with the fusion performance under different weather conditions, different exposure states, and different target scales to ensure that the stitching result remains natural and stable in various scenarios.

[0099] In the above implementation, the simplified background pixel set and the high-priority target region are stitched together using the fusion ratio value to obtain a stitched image matrix. The specific implementation process is as follows: First, the simplified background pixel set and the high-priority target region are mapped to a unified image coordinate system. Then, the stitching boundary and stitching transition zone are determined with the spatial position of the target region as the center. The stitching boundary is the directly adjacent position of the background region and the target region, and the stitching transition zone is the brightness transition range formed by extending to both sides around the stitching boundary. The width of the stitching transition zone is determined based on the edge gradient difference, the target size, and the visual weight of the target in the entire image. When the target size is large or the gradient difference is large, the width of the stitching transition zone is appropriately increased to enhance the smoothness of the connection; when the target size is small and the background structure is simple, the width of the stitching transition zone is appropriately reduced to avoid excessive diffusion of the target boundary. Subsequently, within the stitching transition zone, the background pixels and target pixels are combined point-by-point according to the fusion ratio value, so that the area near the center of the target retains more pixel features of the high-priority target region, and the area near the outer edge of the background inherits more brightness and texture trends of the simplified background pixel set, completing a smooth connection proportionally at the intermediate transition position. For non-transitional areas, the pixel results of the corresponding region are directly used, thus forming a structurally complete and boundary-continuous stitched image matrix across the entire image. After stitching, a continuity check is performed near the stitching boundaries to check for any brightness jumps, edge breaks, or local blur anomalies. If local anomalies are found, the stitching ratio and transition width are fine-tuned within a limited range to ensure that the target area and background area maintain a unified visual appearance in the overall image.

[0100] In the above implementation, a gamma correction algorithm is applied to the stitched image matrix for mapping to obtain the final dynamic range adaptively corrected image. The specific implementation process is as follows: First, the overall brightness distribution in the stitched image matrix is ​​statistically analyzed, including the proportion of low-brightness areas, intermediate brightness areas, and high-brightness areas. Then, combined with the previous global dynamic range analysis results, it is determined whether the current stitched image still suffers from insufficient dark area compression, insufficient bright area suppression, or excessive concentration of intermediate levels. Subsequently, the gamma correction intensity is set based on the statistical results to further achieve a balanced brightness distribution in the overall visual aspect of the image. If the overall stitched image is too dark, the dark area expansion intensity is increased to make the dark details of the background and target areas easier to distinguish; if the overall stitched image is too bright, the bright area suppression effect is enhanced to prevent strong light areas from being overly prominent; if the overall brightness distribution is relatively balanced, the gamma mapping is mainly used to optimize the transition of brightness levels, so that the target area and background area form a more natural visual unity after stitching. The determination of the gamma correction intensity is based on the global brightness statistics of the image, the evaluation results of the stitching boundary continuity, and the requirements for preserving target details. When the intensity is too low, the improvement in overall dynamic range is limited; when the intensity is too high, it may cause the image to appear grayish or local layer compression. Therefore, it is repeatedly adjusted through multiple scene samples to balance brightness balance and structural fidelity. After gamma mapping processing, the final dynamic range adaptively corrected image is obtained. This image maintains clear details in the target area, maintains a natural transition in the background area, and is more suitable for stable recognition by the subsequent visual perception module in terms of overall brightness level.

[0101] S7 includes extracting texture distribution from the dynamic range adaptively corrected image to obtain an initial decision feature set; calculating feature completeness values ​​based on the initial decision feature set, and determining complete decision features or repair decision features based on the feature completeness values; inputting the complete decision features or the repair decision features into the decision module to generate perception control instructions; and performing state mapping based on the perception control instructions to obtain adaptively enhanced visual perception results.

[0102] In the above implementation, the dynamic range adaptive correction image is the final output image after prior region enhancement, edge fusion, and overall brightness adjustment. This image has been optimized in terms of brightness levels, edge structure, and local details. First, texture distribution is extracted from this image to construct an initial decision feature set. The specific implementation process is as follows: the entire image is traversed, dividing it into multiple analysis units. directional change statistics and brightness fluctuation analysis are performed on the pixels within each analysis unit to identify texture-dense and texture-sparse regions. Texture distribution extraction focuses not only on the brightness variation amplitude between pixels but also on the consistency of the direction of change and spatial continuity, thereby distinguishing road surface texture, lane line structure, target contour texture, and background environment texture. The size of the analysis unit is determined based on the image resolution, target scale, and the feature granularity requirements of the subsequent decision module. When the target is small and densely distributed, the analysis unit size is smaller to improve texture recognition accuracy; when the target is large or the scene is simple, the analysis unit size is appropriately increased to reduce computational burden. After completing the texture statistics for each analysis unit, texture intensity, texture orientation consistency, and texture distribution density are mapped to a unified feature space. Combined with edge information and brightness distribution information, an initial decision feature set is formed. This feature set contains both texture information reflecting the target structure and distribution features reflecting the environmental background, providing a foundation for subsequent decision analysis.

[0103] In the above implementation, the feature completeness value is calculated based on the initial decision feature set. The specific implementation process is as follows: First, the various features in the initial decision feature set are classified and statistically analyzed, including target-related texture features, edge continuity features, brightness level features, and spatial distribution features. Then, the coverage and continuity of these features in the entire image are evaluated. For target-related texture features, it is determined whether they form a continuous distribution and cover key areas in the image; for edge continuity features, it is determined whether the main contours are broken or blurred; for brightness level features, it is evaluated whether there is a reasonable transition between different brightness areas; for spatial distribution features, it is checked whether various features are concentrated within a reasonable location range. Through the above multi-dimensional evaluation, the completeness of various features is quantified, and a feature completeness value is obtained. This value is determined by weighting the contribution of each feature to the decision result in the sample image, with features with higher contributions having a larger weight in the completeness calculation. The weight setting is determined by analyzing the impact of different features on target recognition accuracy, path judgment stability, and scene understanding accuracy, so that the feature completeness value can truly reflect the current image's support capability for the decision task.

[0104] In the above implementation, complete decision features or repair decision features are determined based on the feature completeness value. Specifically, the calculated feature completeness value is compared with a preset completeness threshold. When the feature completeness value is higher than the threshold, the current feature set is determined to be a complete decision feature set; when the feature completeness value is lower than the threshold, it is determined to be in a repair state, and repair decision features are generated. The completeness threshold is determined based on the feature completeness statistics of successful and failed recognition scenarios in historical samples. By analyzing the range of feature completeness values ​​when recognition is successful and the range when recognition fails, a boundary value that can effectively distinguish between the two types of situations is selected. This value is then adjusted based on different lighting conditions, different weather environments, and different target density scenarios to ensure that the threshold has stable judgment capabilities under various environments. When repair decision features need to be generated, the missing or discontinuous parts in the initial decision feature set are compensated. Compensation methods include connecting broken edges, moderately enhancing weak texture areas, and smoothing abnormal brightness transition areas, thereby improving the overall feature completeness. The intensity of the repair process is determined based on the degree of feature loss; the higher the degree of loss, the greater the repair intensity. When the degree of loss is low, only local areas are fine-tuned to avoid introducing over-correction.

[0105] In the above implementation, complete decision features or repair decision features are input to the decision module to generate perception control commands. Specifically, the feature set is organized according to the input format of the decision module, ensuring consistency in spatial location, scale, and expression across different feature types, and then transmitted to the decision module. The decision module performs target recognition, path determination, and behavior prediction processing based on the input features, thereby generating corresponding perception control commands. The content of the perception control commands includes vehicle direction adjustment, speed adjustment suggestions, and obstacle avoidance strategies, and their generation process is determined based on the target position, motion trend, and environmental structure information in the feature set. To ensure the stability of the commands, consistency constraints are applied to feature changes in consecutive frames during the generation process, ensuring smooth changes in commands between adjacent time points.

[0106] In the above implementation, an adaptively enhanced visual perception result is obtained by performing state mapping based on the perception control commands. Specifically, the perception control commands are mapped to the current vehicle state space, including vehicle position, driving direction, speed, and environmental perception state. Then, the visual perception result is updated by combining this with the feature distribution in the current image. During the state mapping process, by matching control commands with visual features, the output of the vision system not only reflects the current image information but also the decision-making layer's understanding of the environment, thus forming an adaptively enhanced visual perception result. This result improves spatial positioning accuracy, target recognition stability, and consistency of environmental understanding.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive correction method for image dynamic range in visual perception of new energy vehicles, characterized in that, include: S1. Obtain raw image data from the on-board camera of the new energy vehicle, extract preliminary feature maps through convolutional neural network, determine the boundaries of potential key regions in the image, and obtain feature descriptions including brightness distribution, dynamic range distribution and object contours. S2. Determine the importance of the region based on the feature description. If the brightness distribution is lower than the first preset threshold, it is marked as a low-brightness region. If the brightness distribution is higher than the second preset threshold, it is marked as a high-brightness region. The weight allocation of the low-brightness region or the high-brightness region is enhanced through the attention mechanism to obtain a weighted region priority list. S3. Obtain a subset of pixels in high-priority regions using a weighted region priority list. Process these pixel subsets through adaptive filtering and brightness mapping to determine the optimized detail enhancement image blocks and local dynamic range correction image blocks. S4. Extract edge intensity information from the optimized detail enhancement image patch and the local dynamic range correction image patch, determine whether the edge intensity exceeds the preset threshold, and if it does, fuse it into the overall image frame to obtain the intermediate image with preliminary correction. S5. Calculate the global brightness distribution and dynamic range distribution through the intermediate image, adjust the correction depth of the secondary region according to the brightness distribution and dynamic range distribution, and determine the simplified background pixel set.

2. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 1, characterized in that: S1 includes: A preliminary feature map is obtained by extracting spatial information from the original image matrix using a convolutional neural network. The boundaries of key regions are determined based on the gradient change rate of the preliminary feature map. A brightness distribution histogram is obtained by statistically analyzing the gray-level frequency of the key region boundary. If the extreme value difference of the brightness distribution histogram is greater than a preset threshold, a dynamic range distribution matrix is ​​obtained. By fusing the brightness distribution histogram, the dynamic range distribution matrix, and the object contour lines extracted based on the key region boundaries, a feature description containing brightness distribution, dynamic range distribution, and object contour is obtained.

3. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 1, characterized in that: S2 includes: The initial brightness distribution map is obtained based on the feature description; If the initial brightness distribution map is lower than the first preset threshold, it is marked as a low-brightness area; if it is higher than the second preset threshold, it is marked as a high-brightness area. An attention mechanism is used to perform feature mapping on the low-brightness region and the high-brightness region to obtain a weight allocation vector; The weighted calculation is performed based on the weight allocation vector, and the weighted region priority list is obtained by sorting. The importance of the region is then determined based on the weighted region priority list.

4. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 1, characterized in that: S3 includes: Construct a weighted region priority list based on the initial image blocks, and extract a subset of pixels from the high-priority regions in the weighted region priority list; An adaptive filter is used to process the pixel subset to obtain a smooth pixel subset; A brightness reconstruction pixel subset is obtained by processing the smooth pixel subset using a nonlinear mapping function constructed based on the smooth pixel subset; High-frequency edge information is extracted from the brightness reconstructed pixel subset to determine the optimized detail-enhanced image block, and the brightness reconstructed pixel subset is corrected to determine the local dynamic range corrected image block.

5. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 1, characterized in that: S4 includes: Edge extraction is performed on detail-enhanced image patches and local dynamic range-corrected image patches to obtain local edge intensity information; Determine whether the local edge intensity information exceeds a preset edge intensity threshold; If the local edge intensity information exceeds the edge intensity threshold, then the corresponding pixel is determined to be a high-intensity edge pixel. Obtain the global mapping position of the high-intensity edge pixels in the overall image frame; Based on the global mapping position, the high-intensity edge pixels are frame-fused to obtain a preliminary corrected intermediate image.

6. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 1, characterized in that: S5 includes: An intermediate image is acquired, the intermediate image is processed to obtain a global brightness distribution, and a dynamic range distribution is obtained based on the global brightness distribution. If the contrast of a candidate region in the intermediate image is lower than the threshold set by the dynamic range distribution, the candidate region is determined to be a secondary region. The depth adjustment factor is calculated based on the global brightness distribution and the dynamic range distribution, and the initial depth of the secondary region is updated using the depth adjustment factor to determine the correction depth. The pixels in the secondary region are processed according to the correction depth to determine the simplified background pixel set.

7. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 1, characterized in that, It also includes S6, obtaining the fusion parameters of the simplified background pixel set and the high-priority region, and integrating the two through pixel-level stitching operations to obtain the final dynamic range adaptive correction image, specifically including: Extract the background pixel set and high-priority target region from the initial acquired image, and remove redundant data from the background pixel set to obtain a simplified background pixel set; The fusion ratio is calculated based on the difference in edge gradient between the simplified background pixel set and the high-priority target region.

8. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 7, characterized in that: S6 further includes: The simplified background pixel set and the high-priority target region are spliced ​​together using the fusion ratio value to obtain a spliced ​​image matrix; The stitched image matrix is ​​mapped using a gamma correction algorithm to obtain the final dynamic range adaptive correction image.

9. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 7, characterized in that, It also includes S7, which extracts decision-related features from the final dynamic range adaptively corrected image, judges the feature completeness, and if complete, outputs it to the decision module of the new energy vehicle visual perception system to obtain adaptively enhanced visual perception results, specifically including: The texture distribution is extracted from the dynamic range adaptively corrected image to obtain the initial decision feature set; The feature completeness value is calculated based on the initial decision feature set, and the complete decision feature or the repaired decision feature is determined based on the feature completeness value.

10. The image dynamic range adaptive correction method in visual perception of new energy vehicles according to claim 9, characterized in that: The S7 also includes: The complete decision features or the repair decision features are input into the decision module to generate perception control commands; The state mapping is performed according to the perception control command to obtain an adaptively enhanced visual perception result.

Images