Image enhancement method and system for vehicle-mounted image sensor

Abstract

The present invention relates to the field of image enhancement, and in particular to an image enhancement method and system for an on-board image sensor. The method comprises the following steps: acquiring driving images recorded from multiple angles of a vehicle; performing dynamic scene adaptive exposure adjustment on the driving images recorded from multiple angles of the vehicle, thereby obtaining multiple adaptively exposed driving images; performing deep network noise recognition on the multiple adaptively exposed driving images, and performing dynamic filtering and noise reduction on each sub-image, thereby generating multiple filtered and noise-reduced optimized images; performing image texture analysis on the multiple exposed sub-images, and performing texture distortion reconstruction to construct multiple texture-reconstructed images; performing road object visual recognition on the multiple texture-reconstructed images, and performing adaptive contrast enhancement processing to construct multiple detail-contrast-enhanced images. The present invention achieves high-definition, stable panoramic on-board images.

Description

Technical Field

[0001] The present invention relates to the field of image enhancement technology, and in particular to an image enhancement method and system for a vehicle-mounted image sensor. Background Art

[0002] Automotive image sensors are increasingly used in intelligent driving systems. As a crucial component of vehicle perception systems, they play a vital role. With the continuous advancement of autonomous driving technology and the increasing intelligence of vehicles, automotive image sensors are widely used in various driver assistance systems (ADAS) and autonomous driving systems, such as lane keeping, collision warning, and automatic parking. Automotive image sensors are responsible for collecting visual information of the surrounding environment and providing real-time data to the vehicle's decision-making systems. However, in actual use, automotive image sensors face numerous challenges. Image quality is often severely affected in complex weather and lighting conditions, such as low light, strong sunlight, rain, fog, and snow. This can lead to noise, blur, and low contrast in the images captured by the sensors.

[0003] Traditional in-vehicle image enhancement methods typically rely on hardware design and simple image processing algorithms such as brightness adjustment, contrast enhancement, and denoising. However, these methods often cannot effectively cope with dynamic changes in complex environments, and their processing effectiveness is limited by the hardware performance of the image sensor and environmental interference. Therefore, as the environmental perception requirements of autonomous driving technology continue to increase, traditional image enhancement methods can no longer meet the needs of high-precision and high-robust image analysis. To meet these challenges, the development of intelligent and efficient image enhancement methods to address the various problems encountered by in-vehicle image sensors in practical applications has become a key requirement for improving the performance of in-vehicle image sensors and optimizing the perception capabilities of autonomous driving systems. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes an image enhancement method and system for a vehicle-mounted image sensor to solve at least one of the above technical problems.

[0005] To achieve the above object, the present invention provides an image enhancement method for a vehicle-mounted image sensor, comprising the following steps:

[0006] Step S1: Acquire driving recorded images of the vehicle at multiple angles; perform dynamic scene adaptive exposure adjustment on the driving recorded images of the vehicle at multiple angles, thereby obtaining multiple adaptive exposure driving images;

[0007] Step S2: performing deep network noise recognition on multiple adaptive exposure driving images, and performing dynamic filtering and noise reduction on each sub-image, thereby generating multiple filtered and noise-reduced optimized images;

[0008] Step S3: performing image texture analysis on the multiple exposure sub-images and performing texture distortion reconstruction to construct multiple texture reconstructed images;

[0009] Step S4: performing road object visual recognition on the multiple texture reconstructed images and performing adaptive contrast enhancement processing to construct multiple detail contrast enhanced images;

[0010] Step S5: monitoring the real-time vibration parameters of the vehicle based on the on-board sensors; performing geometric posture vibration distortion analysis and dynamic instantaneous distortion correction processing on the multiple detail contrast enhanced images based on the real-time vibration parameters of the vehicle, thereby generating a dynamic distortion corrected image;

[0011] Step S6: geometrically transform and register the dynamic distortion-corrected images, and perform dynamic panoramic image stitching to construct a panoramic behavior optimization image model.

[0012] The present invention helps to obtain clear driving recording images under different lighting conditions through dynamic scene adaptive exposure adjustment, obtains multiple adaptive exposure driving images to improve image quality, reduces the problem of underexposure or overexposure, and provides better input for subsequent processing. It effectively reduces the noise in the image through deep network noise recognition and dynamic filtering noise reduction, improves image clarity and quality, and generates multiple filtered noise reduction optimized images to help reduce interference and make the image more detailed and clear. Image texture analysis and distortion reconstruction help to restore texture loss caused by noise or other factors in the image, construct multiple texture reconstructed images to enhance the details and texture of the image, improve visual perception effect, and road object visual recognition and contrast enhancement processing help It helps identify objects on the road and enhances detail contrast in images. Constructing multiple detail contrast-enhanced images helps improve image clarity and visual effects, making road objects easier to identify. It reduces image distortion caused by vehicle vibration through geometric posture vibration distortion analysis and dynamic instantaneous distortion correction processing, improves image stability and clarity, and generates dynamic distortion correction images to help eliminate distortion caused by vibration and maintain image accuracy and stability. It seamlessly stitches multiple images into a panoramic image through geometric transformation alignment and dynamic panoramic image stitching, providing a comprehensive perspective and information. Constructing a panoramic behavior optimization image model provides a broader field of view and more comprehensive information, providing a better image reference and analysis basis for the vehicle system.

[0013] Preferably, step S1 includes the following steps:

[0014] Step S11: Acquire driving record images of the vehicle at multiple angles;

[0015] Step S12: performing scene depth semantic segmentation on each of the vehicle's driving record images from multiple angles to obtain scene semantic features of each image;

[0016] Step S13: performing image region segmentation according to the scene semantic features of each image, thereby generating a plurality of different scene semantic regions;

[0017] Step S14: analyzing the dynamic changes of ambient light in a plurality of different scene semantic areas to generate dynamic features of ambient light in different scenes;

[0018] Step S15: performing dynamic scene adaptive exposure adjustment on a plurality of different scene semantic areas according to the dynamic characteristics of ambient light of different scenes, thereby obtaining a plurality of adaptive exposure driving images.

[0019] The present invention provides a comprehensive perspective through driving recording images of the vehicle from multiple angles, contains rich scene information, and provides a rich data source for subsequent processing. Through deep semantic segmentation of image scenes, different objects and scenes in the image are accurately identified, and the scene semantic features of each image are extracted. The scene semantic features are helpful for subsequent image processing and analysis, and improve the accuracy and efficiency of processing. Image region segmentation divides the image into different regions according to the scene semantic features, which helps to further understand and analyze the content in the image. Generating multiple different scene semantic regions provides more refined image information for subsequent processing, which helps to perform more refined processing on different regions. Through dynamic change analysis of ambient light, the lighting changes of different scenes are understood, and the dynamic features of ambient light are extracted. The dynamic features of ambient light help adjust the exposure and contrast of the image to adapt to the image processing requirements under different lighting conditions. Dynamic scene adaptive exposure adjustment is performed according to the dynamic features of ambient light in different scenes. The exposure level of the image is adjusted according to the actual lighting conditions to ensure image quality and visibility. Multiple adaptive exposure driving images are obtained to provide clear and accurate image information under different lighting conditions, which helps to improve the readability and analysis effect of the image.

[0020] Preferably, the dynamic scene adaptive exposure adjustment is specifically:

[0021] Perform scene region classification on multiple different scene semantic regions to obtain sky regions, road regions, vehicle regions, and building regions;

[0022] The dynamic characteristics of ambient light in different scenes specifically include: characteristics of ambient light brightness changes in the sky, characteristics of brightness distribution in the road area, characteristics of vehicle surface light, and local brightness intensity of buildings;

[0023] Based on the brightness variation characteristics of the sky environment, the overexposure reduction processing is performed on the sky area to obtain an overexposure optimized image of the sky area;

[0024] Identify the local brightness distribution of the road area to extract images of dark areas of the road;

[0025] Performing high-exposure processing on the dark road area image according to the brightness distribution characteristics of the road area to generate a high-exposure road area image;

[0026] Identify surface reflective points in the vehicle area to generate reflective feature points on the vehicle surface;

[0027] Performing local exposure adjustment on the reflective feature points on the vehicle surface based on the light features of the vehicle surface to obtain a local exposure vehicle area image;

[0028] Adaptive exposure compensation is performed on the building area based on the local brightness intensity of the building to obtain an adaptive exposure building area image.

[0029] The present invention classifies different scene semantic areas, including sky areas, road areas, vehicle areas and building areas, which helps to understand the content in the image more accurately and provide more accurate information for subsequent processing. These features provide information about the lighting characteristics of different areas, which helps to adjust image parameters according to actual lighting conditions, improve image quality and visibility, and avoid loss of details in the sky area in the image through overexposure reduction processing of the sky area, thereby improving the quality and viewing experience of the overall image. Through local brightness distribution recognition and high exposure processing, details of dark areas of the road are extracted, visibility of these areas is improved, and driving safety is improved. Identifying reflective feature points on the vehicle surface and performing local exposure adjustment helps to highlight vehicle details and improve visibility and texture of the vehicle area. Adaptive exposure compensation is performed according to the local brightness intensity of the building to ensure that the building area is presented in the image more realistically and clearly.

[0030] Preferably, the specific steps of step S2 are:

[0031] Step S21: performing deep network noise recognition on multiple adaptive exposure driving images to extract all noise points in the images;

[0032] Step S22: performing noise distribution analysis on all noise points in the image to generate image noise distribution features;

[0033] Step S23: dividing the image based on the image noise distribution characteristics to obtain multiple exposure sub-images;

[0034] Step S24: calculating the noise distribution level of the multiple exposed sub-images, thereby obtaining the noise distribution level of each sub-image;

[0035] Step S25: performing dynamic filtering and noise reduction on each sub-image based on the noise distribution level of each sub-image, thereby generating a plurality of filtered and noise-reduced optimized images.

[0036] The present invention uses deep network noise recognition to accurately identify noise points in the image, including various types of noise. Effective identification of noise points is helpful for subsequent noise processing and improves image quality and clarity. Noise distribution analysis is performed on all noise points in the image to understand the distribution of noise in the image, providing an important basis for subsequent processing. Noise distribution feature analysis helps to determine the type and degree of noise, providing guidance for subsequent noise reduction processing. Image division is performed based on image noise distribution features to divide the image into different regions and generate multiple exposure sub-images, which helps to independently process the noise in different regions and improve the noise elimination effect. Noise distribution level calculation is performed on multiple exposure sub-images to quantify the degree of noise in each sub-image, helping to determine the appropriate noise reduction strategy. Understanding the noise distribution level helps to select appropriate noise reduction algorithms and parameters and improve the noise reduction effect. Dynamic filtering noise reduction is performed on each sub-image based on the noise distribution level of each sub-image, and different noise reduction strategies are adopted according to the noise conditions in different regions. Dynamic filtering noise reduction helps to retain image details while effectively reducing noise, thereby improving image clarity and quality.

[0037] Preferably, the specific steps of step S3 are:

[0038] Step S31: performing image texture analysis on multiple exposure sub-images to obtain image texture features of different noise backgrounds;

[0039] Step S32: performing multi-scale convolution on the image texture features of different noise backgrounds to generate multi-scale texture convolution features;

[0040] Step S33: performing detailed texture distortion identification on the plurality of filtered noise reduction optimized images, and marking image areas with texture distortion;

[0041] Step S34: performing texture distortion reconstruction on the texture-distorted image region according to the multi-scale texture convolution feature to construct a plurality of texture-reconstructed images.

[0042] The present invention obtains image texture features under different noise backgrounds by performing image texture analysis on multiple exposed sub-images. Texture analysis helps to understand the detailed structure and texture features of the image, providing an important reference for subsequent processing. Multi-scale convolution is performed on the image texture features under different noise backgrounds to generate multi-scale texture convolution features. Multi-scale convolution helps to capture texture information at different scales of the image, improves the description and analysis capabilities of texture features, performs detail texture distortion recognition on multiple filtered denoising optimized images, marks image areas with texture distortion, and effectively identifies detail texture distortion problems in the image through texture distortion recognition, providing optimization direction for subsequent processing. Texture distortion reconstruction is performed on the texture distorted image area according to the multi-scale texture convolution features, and multiple texture reconstructed images are constructed. Texture distortion reconstruction helps to repair texture distortion problems in the image, improving the visual quality and detail performance of the image.

[0043] Preferably, the specific steps of step S4 are:

[0044] Step S41: performing road object visual recognition on the multiple texture reconstructed images and marking multiple road object nodes;

[0045] Step S42: mining road semantic features of multiple road object nodes to obtain road semantic features of each object node;

[0046] Step S43: performing a road content priority analysis based on the road semantic features of each object node to generate a road content priority for each object node;

[0047] Step S44: performing object edge detail mining on multiple road object nodes to generate edge detail features of each object;

[0048] Step S45: Adaptively perform contrast enhancement processing on the edge detail features of each object based on the road content priority of each object node, thereby constructing multiple detail contrast enhanced images.

[0049] The present invention performs road object visual recognition on multiple texture reconstructed images and marks multiple road object nodes, which helps to identify road objects in the image and provides a basis for subsequent analysis and processing. It performs road semantic feature mining on multiple road object nodes to help obtain the road semantic features of each object node. By mining the semantic features, the meaning and attributes of road objects can be better understood. Road content priority analysis is performed based on the road semantic features of each object node to generate the road content priority of each object node. The priority analysis helps to determine the importance and influence of the object in the image and guide the priority of subsequent processing. Object edge detail mining is performed on multiple road object nodes to generate edge detail features of each object. Edge detail mining helps to capture detailed information of the object edge and improve the object recognition. Based on the road content priority, the edge detail features of each object are adaptively contrast enhanced to generate multiple detail contrast enhanced images, highlighting the details of each object in the image and improving the clarity and contrast of the image.

[0050] Preferably, the specific steps of step S5 are:

[0051] Step S51: monitoring the real-time vibration parameters of the vehicle based on the vehicle-mounted sensors;

[0052] Step S52: performing time-series vibration feature evolution on the real-time vibration parameters of the vehicle to obtain the vehicle time-series vibration features;

[0053] Step S53: performing geometric posture vibration distortion analysis on the plurality of detail contrast enhanced images according to the vehicle time-series vibration characteristics, thereby generating a driving image vibration distortion characteristic;

[0054] Step S54: performing dynamic instantaneous distortion correction processing based on the vibration distortion characteristics of the driving image, thereby generating a dynamic distortion corrected image.

[0055] The present invention obtains the real-time vibration parameters of the vehicle by monitoring the on-board sensors, and understands the vibration conditions of the vehicle in real time, which is helpful to evaluate the stability and comfort of the vehicle and provide basic data for subsequent processing. The real-time vibration parameters of the vehicle are subjected to time-series vibration feature evolution to obtain the time-series vibration characteristics of the vehicle. The time-series vibration characteristics are helpful to analyze the laws and characteristics of vehicle vibration and provide a basis for vibration characteristics for subsequent processing. According to the time-series vibration characteristics of the vehicle, geometric posture vibration distortion analysis is performed on multiple detail contrast enhanced images to generate driving image vibration distortion characteristics, which is helpful to understand the impact of vehicle vibration on the image and provide a basis for vibration correction for subsequent processing. Dynamic instantaneous distortion correction processing is performed based on the vibration distortion characteristics of the driving image to generate a dynamic distortion corrected image. Through the distortion correction processing, the image distortion caused by vehicle vibration is eliminated and the clarity and stability of the image are improved.

[0056] Preferably, the specific steps of step S6 are:

[0057] Step S61: segmenting the dynamic distortion corrected image into continuous frames, thereby obtaining continuous frame corrected images at multiple angles;

[0058] Step S62: performing multi-angle image feature point recognition on the continuous frame corrected images at multiple angles, and extracting the image continuous frame feature points at each angle;

[0059] Step S63: performing geometric transformation registration on the feature points of the continuous frames of the images at each angle, thereby generating multi-frame registration data of the feature points;

[0060] Step S64: performing dynamic panoramic image stitching on the dynamic distortion corrected image based on the multi-frame registration data of the feature points, thereby generating a dynamic panoramic stitching image;

[0061] Step S65: performing local smoothing processing on the spliced ​​portions of the dynamic panoramic spliced ​​image, thereby constructing a panoramic behavior optimized image model.

[0062] The present invention helps to separate the continuous frame images into separate frames by performing continuous frame segmentation on the dynamic distortion correction image, so that each frame is processed and analyzed separately, and the continuous frame correction images of multiple angles are obtained to provide more angle viewing information, which provides more flexibility for subsequent processing and analysis. By multi-angle image feature point recognition, the continuous frame feature points of each angle are extracted for subsequent image registration and splicing. The recognition of feature points helps to capture the key information in the image and help improve the accuracy and efficiency of image registration. The geometric transformation registration is performed on the continuous frame feature points of each angle of the image, and the images between different frames are aligned to improve the consistency and coherence of the image and generate feature points. The multi-frame registration data of the points helps to eliminate the offset and distortion between images, and provides an accurate data basis for subsequent image stitching. The dynamic distortion-corrected images are dynamically stitched based on the multi-frame registration data of the feature points, and the continuous frame-corrected images at different angles are stitched into a panoramic image. The dynamic panoramic stitched image provides a wider field of view and more comprehensive information, providing a better image reference and analysis basis for the vehicle-mounted system. Local smoothing of the spliced ​​parts of the dynamic panoramic stitched image helps to eliminate the discontinuity at the spliced ​​parts, making the panoramic image smoother and continuous. The panoramic behavior optimization image model is constructed to provide more realistic and accurate panoramic images, and improve the visual effect and usability of the image.

[0063] In this specification, an image enhancement system for a vehicle-mounted image sensor is provided, which is used to perform the image enhancement method for the vehicle-mounted image sensor as described above, including:

[0064] The adaptive exposure module is used to obtain driving recording images of the vehicle at multiple angles; the driving recording images of the vehicle at multiple angles are subjected to dynamic scene adaptive exposure adjustment, thereby obtaining multiple adaptive exposure driving images;

[0065] The filtering and denoising module is used to perform deep network noise recognition on multiple adaptive exposure driving images and perform dynamic filtering and denoising on each sub-image, thereby generating multiple filtered and denoised optimized images;

[0066] A texture reconstruction module is used to perform image texture analysis on multiple exposure sub-images and perform texture distortion reconstruction to construct multiple texture reconstructed images;

[0067] A contrast enhancement module is used to perform visual recognition of road objects on multiple texture-reconstructed images and perform adaptive contrast enhancement processing to construct multiple detail contrast-enhanced images;

[0068] The distortion correction module is used to monitor the real-time vibration parameters of the vehicle based on on-board sensors. Based on the real-time vibration parameters of the vehicle, the module performs geometric posture vibration distortion analysis and dynamic instantaneous distortion correction on multiple detail contrast-enhanced images to generate dynamic distortion-corrected images.

[0069] The panoramic stitching module is used to perform geometric transformation and registration on the dynamic distortion correction images and perform dynamic panoramic image stitching to construct a panoramic behavior optimization image model.

[0070] The present invention ensures the acquisition of high-quality driving recording images under different lighting conditions through dynamic scene adaptive exposure adjustment. The adaptive exposure driving images at multiple angles help to improve the brightness and contrast of the image, enhance the details and information content of the image, and combine deep network noise recognition with dynamic filtering noise reduction to effectively remove noise in the image, improve the clarity and quality of the image. The generated filtered noise reduction optimized image reduces noise and artifacts in the image, improves the visual effect and recognition accuracy of the image, and improves the details and texture of the image through texture distortion reconstruction, making the image more realistic and clear. The constructed texture reconstruction image reduces blur and distortion in the image, enhances the texture characteristics of the image, and improves the recognition and analysis capabilities of the image. The adaptive contrast enhancement processing is effective. It helps to highlight the details and edges in the image, enhance the contrast and visual effects of the image. The generated detail contrast enhanced image improves the readability and recognition of the image, making road objects more clearly visible. It performs distortion correction processing based on real-time vibration parameters to effectively eliminate the distortion in the image caused by vehicle vibration and improve the accuracy and stability of the image. The dynamic distortion correction image can provide a more realistic and clear perspective, providing a more reliable data basis for subsequent analysis and processing. The dynamic distortion correction image is subjected to geometric transformation alignment and panoramic stitching to generate a panoramic behavior optimization image model, providing a wider field of view and more complete information. The constructed panoramic behavior optimization image model helps the vehicle system better understand the surrounding environment and improve driving safety and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic flow chart of the steps of an image enhancement method for a vehicle-mounted image sensor according to the present invention;

[0072] Figure 2 Detailed implementation flow chart of step S1;

[0073] Figure 3 Detailed implementation flow chart of step S2;

[0074] Figure 4 Schematic diagram of the detailed implementation steps of step S3. DETAILED DESCRIPTION

[0075] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0076] This application provides an image enhancement method and system for an on-board image sensor. The execution entities of the image enhancement method and system for an on-board image sensor include, but are not limited to, the following: mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered as general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio and image management system, an information management system, and a cloud data management system.

[0077] See also Figures 1 to 4 The present invention provides an image enhancement method for a vehicle-mounted image sensor, the image enhancement method for the vehicle-mounted image sensor comprising the following steps:

[0078] Step S1: Acquire driving recorded images of the vehicle at multiple angles; perform dynamic scene adaptive exposure adjustment on the driving recorded images of the vehicle at multiple angles, thereby obtaining multiple adaptive exposure driving images;

[0079] Step S2: performing deep network noise recognition on multiple adaptive exposure driving images, and performing dynamic filtering and noise reduction on each sub-image, thereby generating multiple filtered and noise-reduced optimized images;

[0080] Step S3: performing image texture analysis on the multiple exposure sub-images and performing texture distortion reconstruction to construct multiple texture reconstructed images;

[0081] Step S4: performing road object visual recognition on the multiple texture reconstructed images and performing adaptive contrast enhancement processing to construct multiple detail contrast enhanced images;

[0082] Step S5: monitoring the real-time vibration parameters of the vehicle based on the on-board sensors; performing geometric posture vibration distortion analysis and dynamic instantaneous distortion correction processing on the multiple detail contrast enhanced images based on the real-time vibration parameters of the vehicle, thereby generating a dynamic distortion corrected image;

[0083] Step S6: geometrically transform and register the dynamic distortion-corrected images, and perform dynamic panoramic image stitching to construct a panoramic behavior optimization image model.

[0084] The present invention helps to obtain clear driving recording images under different lighting conditions through dynamic scene adaptive exposure adjustment, obtains multiple adaptive exposure driving images to improve image quality, reduces the problem of underexposure or overexposure, and provides better input for subsequent processing. It effectively reduces the noise in the image through deep network noise recognition and dynamic filtering noise reduction, improves image clarity and quality, and generates multiple filtered noise reduction optimized images to help reduce interference and make the image more detailed and clear. Image texture analysis and distortion reconstruction help to restore texture loss caused by noise or other factors in the image, construct multiple texture reconstructed images to enhance the details and texture of the image, improve visual perception effect, and road object visual recognition and contrast enhancement processing help It helps identify objects on the road and enhances detail contrast in images. Constructing multiple detail contrast-enhanced images helps improve image clarity and visual effects, making road objects easier to identify. It reduces image distortion caused by vehicle vibration through geometric posture vibration distortion analysis and dynamic instantaneous distortion correction processing, improves image stability and clarity, and generates dynamic distortion correction images to help eliminate distortion caused by vibration and maintain image accuracy and stability. It seamlessly stitches multiple images into a panoramic image through geometric transformation alignment and dynamic panoramic image stitching, providing a comprehensive perspective and information. Constructing a panoramic behavior optimization image model provides a broader field of view and more comprehensive information, providing a better image reference and analysis basis for the vehicle system.

[0085] In the embodiment of the present invention, see Figure 1 , is a schematic flow chart of the steps of an image enhancement method for a vehicle-mounted image sensor according to the present invention. In this example, the steps of the image enhancement method for a vehicle-mounted image sensor include:

[0086] Step S1: Acquire driving recorded images of the vehicle at multiple angles; perform dynamic scene adaptive exposure adjustment on the driving recorded images of the vehicle at multiple angles, thereby obtaining multiple adaptive exposure driving images;

[0087] In this embodiment, a suitable camera device is selected, such as an on-board camera (e.g., a panoramic camera, a single-lens or dual-lens camera), ensuring that it has high-definition and wide-angle shooting capabilities to capture dynamic scenes around the vehicle. Cameras are installed at different locations of the vehicle (e.g., front, side, and rear) to ensure that all important angles around the vehicle are covered. This arrangement helps to obtain comprehensive driving record images, which is convenient for subsequent processing. Different driving scenes are selected for data collection, including urban roads, highways, night driving, and complex weather conditions (e.g., rainy days, snowy days), etc. These diverse scenes will provide rich data sets to help with subsequent adaptive exposure adjustment. During the vehicle's driving process, driving record images from multiple angles are recorded in real time, and a fixed frame rate (e.g., 30 frames per second) is set to ensure the continuity and smoothness of data collection. All recorded data should be saved in a high-quality video format for subsequent processing. The collected driving record images are stored in the vehicle's central control system or cloud storage to ensure the integrity and accessibility of the data for subsequent processing and analysis. The stored data is classified and labeled, including timestamps, geographic location information, etc. The image processing system uses the image processing and analysis to analyze the scene information and scene type, and selects an appropriate adaptive exposure adjustment algorithm, such as histogram equalization, local contrast enhancement, or deep learning-based image enhancement methods. These algorithms can dynamically adjust the image exposure level according to the scene's lighting conditions. Each driving record image is analyzed to evaluate its brightness and contrast. A histogram analysis method is used to calculate the distribution of pixels at different brightness levels in the image to determine the image exposure requirements. Based on the analysis results, the image exposure settings are dynamically adjusted. For example, in low-light conditions, the exposure time or ISO value is increased; in over-light conditions, the exposure time or ISO value is reduced. This process should be real-time to ensure rapid response to changing lighting conditions. The adjusted exposure parameters are applied to the original driving record image to generate multiple adaptive exposure driving records. These images should preserve details while minimizing noise and distortion to ensure improved image quality. The generated adaptive exposure images are compared with the original images to evaluate the improvements in brightness and contrast. The parameters used in the adjustment process are recorded to facilitate subsequent analysis and optimization.

[0088] Step S2: performing deep network noise recognition on multiple adaptive exposure driving images, and performing dynamic filtering and noise reduction on each sub-image, thereby generating multiple filtered and noise-reduced optimized images;

[0089] In this embodiment, a dataset of images to be processed is prepared from the adaptive exposure driving images acquired in the previous stage. These images should cover a variety of scenes and lighting conditions to ensure the diversity and robustness of the model. Image samples containing different types of noise (such as Gaussian noise, salt and pepper noise, etc.) are collected and annotated. These samples will be used to train a deep learning model to improve the accuracy of noise recognition. A suitable deep learning model is selected for noise recognition. Commonly used models include convolutional neural networks (CNN) or more advanced architectures such as U-Net or ResNet. When selecting a model, its performance in image processing tasks should be considered. The selected deep learning model is trained using the prepared noise sample dataset. During the training process, the cross entropy loss function or the mean squared error loss function is used for optimization to improve the noise recognition accuracy of the model. The model performance is evaluated using an independent validation set. The accuracy, recall rate, and F1 score are calculated to ensure that the model can accurately recognize different types of noise. The adaptive exposure driving images to be processed are input into the trained noise recognition model one by one to identify the noise type. The model will output the noise type and corresponding noise of each image. Intensity information is recorded for each image, including the noise type and intensity. This information will be used for subsequent dynamic filtering and noise reduction processing. Based on the noise identification results, an appropriate dynamic filtering algorithm is selected for noise reduction. For example, for Gaussian noise, a Gaussian filter is selected; for salt and pepper noise, a median filter is selected. For each image, the filter parameters are dynamically adjusted based on the noise identification results. For example, if high-intensity Gaussian noise is identified, the filter standard deviation is increased to enhance the noise reduction effect; if salt and pepper noise is identified, the median filter window size is adjusted. The selected dynamic filtering algorithm is applied to filter and denoise the adaptive exposure driving image, generating multiple filtered and denoised optimized images to ensure that image detail and quality are preserved after denoising while reducing the impact of noise. The filtered and denoised optimized images are evaluated through visual evaluation and quantitative indicators (such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM)) to ensure that the image quality after denoising is significantly improved. The processed filtered and denoised optimized images are saved to a database, and the parameter settings and effect evaluation results of the filtering process are recorded for subsequent analysis and processing.

[0090] Step S3: performing image texture analysis on the multiple exposure sub-images and performing texture distortion reconstruction to construct multiple texture reconstructed images;

[0091] In this embodiment, multiple exposure sub-images are selected from the filtered noise reduction optimized image obtained in the previous stage. These sub-images should contain rich texture information for subsequent texture analysis and reconstruction. The texture features that need to be analyzed are determined, including texture direction, frequency, roughness, and contrast. Gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), or other texture feature extraction methods are used to quantify the texture information in the image. A suitable texture feature extraction algorithm is selected. For example, GLCM is used to calculate image features such as contrast, entropy, correlation, and uniformity. The LBP method is used to extract local texture features. The texture feature extraction algorithm is executed on each exposure sub-image to generate feature vectors. These feature vectors will be used in the subsequent texture distortion reconstruction process. The extracted texture features are recorded in a database to ensure that the features of each sub-image are saved for subsequent use. Texture distortion analysis is performed on each exposure sub-image to identify the type of distortion in the image (such as blur, noise, missing, etc.). Image quality assessment indicators (such as PSNR and SSIM) are used to quantitatively evaluate the degree of image distortion. Based on the identified distortion type, a corresponding distortion model is established. For example, for blur distortion, a blur kernel model is used; for noise distortion, a noise model is established. These models will provide the basis for subsequent texture reconstruction. Appropriate texture reconstruction algorithms are selected, such as image reconstruction methods based on convolutional neural networks (CNNs), interpolation methods, or optimization algorithms. These methods can effectively handle texture distortion in images. For each exposed sub-image, texture distortion reconstruction is performed using the established distortion model and the selected reconstruction algorithm. The reconstruction parameters are dynamically adjusted according to the extracted texture features and distortion type to obtain the best effect. Multiple texture reconstructed images are generated to ensure that the image details and texture features are retained as much as possible during the reconstruction process. The reconstructed image should have improved visual quality and texture information. The texture reconstructed image is evaluated through visual evaluation and quantitative indicators (such as PSNR and SSIM) to ensure that the reconstructed image has significant improvements in texture details and overall quality. The processed texture reconstructed image is saved to the database, and the parameter settings and effect evaluation results during the reconstruction process are recorded for subsequent analysis and processing.

[0092] Step S4: performing road object visual recognition on the multiple texture reconstructed images and performing adaptive contrast enhancement processing to construct multiple detail contrast enhanced images;

[0093] In this embodiment, a dataset to be processed is prepared from multiple texture reconstructed images obtained in the previous stage. These images should cover different road scenes and objects to ensure the diversity and effectiveness of the recognition process. The categories of road objects that need to be recognized, such as cars, pedestrians, traffic signs, and obstacles, are determined. A clear target definition will be helpful for subsequent model training and evaluation. A suitable deep learning model is selected for road object visual recognition. Commonly used models include YOLO (You Only Look Once), Faster R-CNN, or SSD (Single Shot Multi Box Detector). These models can recognize multiple objects in an image in real time and accurately. The selected model is trained using a training set containing labeled data. The labeled data should contain the location information and category labels of multiple road objects. During the training process, the cross entropy loss function is used for optimization to improve recognition accuracy. The performance of the trained model is evaluated on an independent validation set. The accuracy, recall, and F1 score are calculated to ensure that the model can accurately recognize objects. To identify the target object, each texture reconstructed image is input into the trained object recognition model for inference. The model will analyze the image and mark all recognized road objects, provide the category, location and confidence information of each object, and record the recognition results of each object, including category, location information and its confidence. This information will be used for subsequent contrast enhancement processing. A suitable adaptive contrast enhancement algorithm is selected, such as adaptive histogram equalization (CLAHE) or gamma adjustment. These algorithms can dynamically adjust the contrast according to the distribution of objects in the image and enhance image details. For each texture reconstructed image, the contrast enhancement parameters are dynamically adjusted according to the recognition results of the object. For example: for high-priority objects (such as pedestrians and traffic lights) identified, the contrast of their area is increased; for background areas, the contrast is appropriately reduced to highlight foreground objects. The selected adaptive contrast enhancement algorithm is applied to each texture reconstructed image to generate multiple detail contrast enhanced images to ensure that the enhanced image is clearer in terms of object edges and details while reducing the impact of noise.

[0094] Step S5: monitoring the real-time vibration parameters of the vehicle based on the on-board sensors; performing geometric posture vibration distortion analysis and dynamic instantaneous distortion correction processing on the multiple detail contrast enhanced images based on the real-time vibration parameters of the vehicle, thereby generating a dynamic distortion corrected image;

[0095] In this embodiment, suitable vehicle-mounted sensors, such as accelerometers and gyroscopes, are selected to monitor the real-time vibration parameters of the vehicle during driving. These sensors should be installed on the chassis, wheels or body of the vehicle to ensure that various vibration data can be accurately captured. A real-time data acquisition system is developed to transmit the vibration data collected by the sensors to the central processing unit through wireless communication technology (such as Bluetooth or Wi-Fi) to ensure that the data can be recorded in real time, including information such as timestamps, vibration amplitudes and frequencies. The collected vibration data is stored in a database for classification and management for subsequent vibration analysis and correction processing. These data will provide an important reference for subsequent geometric posture vibration distortion analysis. Appropriate geometric distortion analysis methods are selected, such as affine transformation, perspective transformation or methods based on feature point matching. These methods can effectively analyze the geometric distortion caused by vehicle vibration in the image and compare the collected real-time vibration parameters with multiple details. The vibration characteristics are combined with the enhanced image to analyze the distortion caused by vibration in the image. By mapping the vibration characteristics to the image characteristics, the affected area is identified and the type and degree of distortion are determined. Based on the analysis results, a geometric distortion model is constructed to describe the deformation caused by vibration in the image. These models will provide a basis for subsequent dynamic instantaneous distortion correction. Appropriate dynamic instantaneous distortion correction algorithms are selected, such as interpolation-based methods, optical flow methods or deep learning-based image reconstruction technology. These algorithms can perform effective image correction based on real-time vibration data and distortion models. For each detail contrast enhanced image, dynamic instantaneous distortion correction is performed using the established distortion model and the selected correction algorithm. By adjusting the correction parameters in real time, it is ensured that the correction can adapt to different vibration states. After correction processing, a dynamic distortion corrected image is generated to ensure that the corrected image is consistent with the original image in terms of geometry and details, while effectively eliminating the distortion caused by vibration.

[0096] Step S6: geometrically transform and register the dynamic distortion-corrected images, and perform dynamic panoramic image stitching to construct a panoramic behavior optimization image model.

[0097] In this embodiment, a sequence of images to be processed is selected from the dynamic distortion corrected images generated in the previous stage. These images should cover different perspectives and time periods to facilitate the subsequent stitching process. The collected images are preprocessed, including denoising, color correction, and brightness adjustment. These preprocessing steps help improve the subsequent registration and stitching effects and ensure that the images are consistent in color and brightness. A suitable geometric transformation registration algorithm is selected, such as RANSAC (random sampling consensus algorithm) based on feature point matching or a method based on optical flow. These algorithms can effectively process geometric transformations between images to achieve accurate registration effects. Feature points are extracted from the dynamic distortion corrected images. Key feature points in each image are extracted using algorithms such as SIFT, ORB, or SURF. Then, the feature points are matched using the nearest neighbor matching or FLANN matching algorithm to obtain corresponding points between each pair of images. The matched feature points are screened using the RANSAC algorithm to estimate the geometric transformation matrix between each pair of images. This matrix will be used in the subsequent image registration process to ensure the accuracy and robustness of the registration. Based on the estimated geometric transformation matrix, the dynamic distortion corrected images are Transformation is performed to complete the image registration process. Interpolation methods (such as bilinear interpolation or cubic interpolation) are used to process the pixel values ​​of the images to ensure the quality of the registered images. An appropriate stitching algorithm is selected, such as image stitching technology based on multi-view geometry or image fusion methods. Common algorithms include Poisson image editing, uniformly weighted averaging, or pyramid-based fusion algorithms. The registered images are stitched together. Based on the registration results of each image, their positions in the panorama are determined and fused. During the stitching process, overlapping areas are seamlessly processed to reduce the visibility of seams. The final dynamic panoramic image is generated, ensuring that the stitched image is visually coherent, with clear edges of all objects and no obvious stitching artifacts. Based on the generated dynamic panoramic image, a panoramic behavior optimization image model is designed. This model should consider aspects such as image feature extraction, behavior analysis, and dynamic scene understanding to improve the effectiveness of subsequent intelligent analysis and decision-making. The optimization model is trained using machine learning or deep learning algorithms, and its performance in specific tasks (such as object detection and behavior recognition) is evaluated. The model is continuously optimized based on the evaluation results to ensure its effectiveness in practical applications.

[0098] In this embodiment, refer to Figure 2 , is a flowchart of the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include:

[0099] Step S11: Acquire driving record images of the vehicle at multiple angles;

[0100] Step S12: performing scene depth semantic segmentation on each of the vehicle's driving record images from multiple angles to obtain scene semantic features of each image;

[0101] Step S13: performing image region segmentation according to the scene semantic features of each image, thereby generating a plurality of different scene semantic regions;

[0102] Step S14: analyzing the dynamic changes of ambient light in a plurality of different scene semantic areas to generate dynamic features of ambient light in different scenes;

[0103] Step S15: performing dynamic scene adaptive exposure adjustment on a plurality of different scene semantic areas according to the dynamic characteristics of ambient light of different scenes, thereby obtaining a plurality of adaptive exposure driving images.

[0104] In this embodiment, a high-resolution camera is selected and equipped with a wide-angle lens to ensure that a wide field of view around the vehicle can be captured. Multiple cameras are used to achieve multi-angle synchronous shooting to ensure coverage of different perspectives (front view, rear view, side view, etc.) and to ensure shooting in different driving environments, including urban roads, highways, and rural roads, to obtain diverse image data. A suitable shooting time is set to capture scenes under different lighting conditions. During the driving process of the vehicle, the on-board computer is used to control the camera for real-time image acquisition and record driving images from multiple angles, ensuring that images from each angle are synchronously acquired within the same time period for subsequent analysis. Based on the semantic segmentation results of each image, a region segmentation algorithm (such as connected region labeling, graph cut, etc.) is applied to extract different scene semantic regions. Each region should correspond to a part of the image with similar semantic features. The features of each segmented region, including area, shape, position, etc., are calculated, and this information is marked for subsequent analysis. The extracted semantic regions and their feature information are stored in a database for subsequent dynamic lighting. Analysis and exposure adjustment: For each scene semantic area, use light sensors or image processing technology to measure the ambient light intensity, use the brightness information of the image as an indicator of the lighting conditions, collect lighting data in different time periods, perform time series analysis on it, and identify the dynamic change pattern of lighting. This is achieved through statistical analysis or machine learning models (such as time series prediction models). Key dynamic characteristics of ambient light are extracted from the analysis results, including light intensity fluctuations, change rate and periodicity, etc., to provide a basis for subsequent scene adaptive exposure. An adaptive exposure control algorithm is designed to determine the optimal exposure setting based on the dynamic characteristics of ambient light in different scenes. Fuzzy logic control or rule-based control methods are used. During the image acquisition process, the camera exposure time, aperture and ISO value are dynamically adjusted according to the real-time monitored lighting conditions to ensure the clarity and detail of the image under different lighting conditions. The adjusted images are output to generate multiple adaptive exposure driving images. These images should be able to clearly reflect the details of different scenes and provide high-quality visual data for subsequent analysis and processing.

[0105] In this embodiment, the dynamic scene adaptive exposure adjustment is specifically as follows:

[0106] Perform scene region classification on multiple different scene semantic regions to obtain sky regions, road regions, vehicle regions, and building regions;

[0107] The dynamic characteristics of ambient light in different scenes specifically include: characteristics of ambient light brightness changes in the sky, characteristics of brightness distribution in the road area, characteristics of vehicle surface light, and local brightness intensity of buildings;

[0108] Based on the brightness variation characteristics of the sky environment, the overexposure reduction processing is performed on the sky area to obtain an overexposure optimized image of the sky area;

[0109] Identify the local brightness distribution of the road area to extract images of dark areas of the road;

[0110] Performing high-exposure processing on the dark road area image according to the brightness distribution characteristics of the road area to generate a high-exposure road area image;

[0111] Identify surface reflective points in the vehicle area to generate reflective feature points on the vehicle surface;

[0112] Performing local exposure adjustment on the reflective feature points on the vehicle surface based on the light features of the vehicle surface to obtain a local exposure vehicle area image;

[0113] Adaptive exposure compensation is performed on the building area based on the local brightness intensity of the building to obtain an adaptive exposure building area image.

[0114] In this embodiment, images of multiple different scenes are collected to ensure that these images contain different environmental features, such as the sky, roads, vehicles, and buildings. The data set should cover various lighting conditions and weather conditions to enhance the generalization ability of the model. Deep learning technology is adopted, and convolutional neural networks (CNN) or more advanced network structures (such as U-Net or MaskR-CNN) are used to perform semantic segmentation of images. Annotated data sets are prepared, including manual annotation of each scene area, to facilitate model training. The trained model is used to process the input driving record images frame by frame, and different scene semantic regions in each image are extracted. The model segments the image into Generate a binary mask for each area of ​​the sky, road, vehicle and building areas, analyze the light intensity changes in the sky area, and capture the light changes by calculating the mean and standard deviation of the light intensity. Road area brightness distribution characteristics: Analyze the brightness distribution of the road area through the histogram to identify the concentrated area of ​​brightness and the dark area. Vehicle surface light characteristics: Extract the light reflection characteristics of the vehicle area, identify the reflection point and glossiness, local brightness intensity of the building: Perform local brightness analysis on the building area to determine the brightness intensity difference of different parts. According to the brightness change characteristics of the sky environment light, perform overexposure detection on the sky area and set a threshold. When the brightness When the threshold is exceeded, it is judged as overexposure. Image processing algorithms (such as histogram equalization or gamma correction) are used to optimize the overexposed sky area and reduce the brightness of the overexposed part. Software tools (such as OpenCV) are used to make pixel-level adjustments to obtain an overexposed optimized image of the sky area. Based on the brightness distribution characteristics of the road area, dark areas are identified. By setting a brightness threshold, areas below a specific brightness are found. High exposure adjustment is applied to the identified dark road areas. Weighted averaging or local contrast enhancement technology is used to increase the brightness of these areas, thereby generating a high-exposure road area image. Image processing techniques (such as edge detection or threshold segmentation) are used to ), identify surface reflective points in the vehicle area, which usually have high brightness and form a sharp contrast with the surrounding area, perform local exposure adjustment on the identified reflective feature points, reduce the brightness of the reflective points, avoid overexposure, and use a local brightness adjustment algorithm to ensure that the details of the vehicle surface are preserved. Based on the local brightness intensity data of the building, identify the areas that need exposure compensation, analyze the brightness characteristics of different parts, set the compensation strategy, perform adaptive exposure compensation on the building area, use local contrast enhancement technology, adjust the exposure value, and ensure that the details of the building are clearly visible. This process can be implemented using an image processing library (such as OpenCV).

[0115] In this embodiment, refer to Figure 3 , is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include:

[0116] Step S21: performing deep network noise recognition on multiple adaptive exposure driving images to extract all noise points in the images;

[0117] Step S22: performing noise distribution analysis on all noise points in the image to generate image noise distribution features;

[0118] Step S23: dividing the image based on the image noise distribution characteristics to obtain multiple exposure sub-images;

[0119] Step S24: calculating the noise distribution level of the multiple exposed sub-images, thereby obtaining the noise distribution level of each sub-image;

[0120] Step S25: performing dynamic filtering and noise reduction on each sub-image based on the noise distribution level of each sub-image, thereby generating a plurality of filtered and noise-reduced optimized images.

[0121] In this embodiment, the driving image data collected from the previous adaptive exposure processing step should contain different lighting conditions and exposure settings to ensure comprehensive noise recognition. A suitable deep learning model is selected for noise recognition, and a convolutional neural network (CNN) is used for image noise detection. Alternatively, a pre-trained deep learning model (such as U-Net or ResNet) is used for image segmentation to specifically identify noise points. The model is trained using the labeled noise data as a training set. After training, the adaptive exposure driving image is input into the model to extract all noise in the image. Points, the process should generate a binary mask, mark the noise area identified in the image, extract the coordinates and intensity information of all noise points in the image, perform statistical analysis on the extracted noise points, calculate the intensity mean, variance and spatial distribution of the noise in the image, use the histogram analysis method to generate a histogram of the noise intensity to identify the main distribution area of ​​the noise, and record the noise distribution characteristics in a feature vector, including the mean, standard deviation, maximum and minimum values ​​of the noise intensity, to provide a basis for subsequent image segmentation and noise reduction processing. According to the noise distribution characteristics, formulate an image segmentation strategy and select Select a region-based partitioning method, such as dividing the image into several fixed-size sub-image blocks or performing adaptive partitioning based on changes in noise intensity. According to the defined partitioning strategy, the original driving image is segmented to generate multiple exposure sub-images. Each sub-image should retain its original noise characteristics for subsequent analysis. A noise distribution level calculation method is defined, using statistical characteristics such as the mean and standard deviation of noise intensity as noise level indicators. The noise distribution level is calculated for each exposure sub-image, and the statistical characteristics of noise intensity are extracted and recorded. A noise distribution level report for each sub-image is generated, including information such as the noise mean and variance. Based on the noise distribution level, an appropriate filter is selected for each sub-image, using a mean filter, a Gaussian filter, or a more complex non-local means (NLM) filter. The filter parameters are dynamically adjusted based on the noise level. The selected dynamic filter is applied to each exposure sub-image for noise reduction. The filter kernel size and strength are adjusted based on the noise level to achieve the best noise reduction effect. Multiple filtered noise reduction optimized images are generated. These images should effectively reduce noise while preserving details. The optimized images are saved and the processing parameters are recorded for subsequent analysis and application.

[0122] In this embodiment, refer to Figure 4 , is a flowchart of the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include:

[0123] Step S31: performing image texture analysis on multiple exposure sub-images to obtain image texture features of different noise backgrounds;

[0124] Step S32: performing multi-scale convolution on the image texture features of different noise backgrounds to generate multi-scale texture convolution features;

[0125] Step S33: performing detailed texture distortion identification on the plurality of filtered noise reduction optimized images, and marking image areas with texture distortion;

[0126] Step S34: reconstructing the texture-distorted image region according to the multi-scale texture convolution feature to construct a plurality of texture-reconstructed images.

[0127] In this embodiment, multiple exposure sub-images that have undergone filtering and noise reduction processing are obtained from the previous steps. These images should contain different noise backgrounds to ensure data diversity. Texture analysis methods such as gray-level co-occurrence matrix (GLCM), local binary pattern (LBP) or wavelet transform are used to extract the texture features of the image. These methods can effectively capture the texture information under different noise backgrounds in the image. Texture features are extracted for each exposure sub-image, and the selected texture analysis algorithm is used to generate corresponding texture feature vectors. These features include indicators such as contrast, correlation, energy and homogeneity, which describe the texture of different areas in the image. Processing information, design multi-scale convolution kernels, usually use convolution kernels of different sizes (such as 3x3, 5x5, 7x7) to capture the details and texture features in the image, use Gaussian blur convolution kernel, edge detection convolution kernel, etc., perform multi-scale convolution operation on the texture features of each exposed sub-image, apply the designed convolution kernel to the texture feature map to extract texture information at different scales, generate multi-scale texture convolution feature map through convolution operation, which can more comprehensively reflect the texture characteristics of the image, integrate the convolution features of different scales into a feature vector, and provide support for subsequent texture distortion recognition, using feature splicing or weighting The averaging method integrates features of different scales, selects a suitable algorithm for texture distortion recognition, and uses machine learning-based algorithms such as support vector machines (SVM) or random forests, or deep learning-based methods such as convolutional neural networks (CNN) to mark the detailed distortion in the image. The optimized image after filtering and noise reduction is input into the distortion recognition model for texture distortion detection. The model will analyze the texture features of the image, identify the distorted areas, and generate corresponding masks to mark the distorted areas. Based on the marking results, the location information of the distorted areas in each image is recorded in preparation for subsequent texture reconstruction processing. Provide a basis for selecting a suitable texture reconstruction algorithm, such as texture synthesis method or image restoration technology, and use neighborhood-based methods (such as Patch-based synthesis) or deep learning image restoration networks (such as GAN) for reconstruction. For each marked texture distortion area, apply the selected reconstruction algorithm and use multi-scale texture convolution features to generate the corresponding texture reconstructed image, restore the distorted details and improve the image quality. Save the texture reconstructed image and record the parameter settings and effect evaluation during the reconstruction process. These reconstructed images should reflect clearer texture details to provide support for subsequent applications.

[0128] In this embodiment, step S4 includes the following steps:

[0129] Step S41: performing road object visual recognition on the multiple texture reconstructed images and marking multiple road object nodes;

[0130] Step S42: mining road semantic features of multiple road object nodes to obtain road semantic features of each object node;

[0131] Step S43: performing a road content priority analysis based on the road semantic features of each object node to generate a road content priority for each object node;

[0132] Step S44: performing object edge detail mining on multiple road object nodes to generate edge detail features of each object;

[0133] Step S45: Adaptively perform contrast enhancement processing on the edge detail features of each object based on the road content priority of each object node, thereby constructing multiple detail contrast enhanced images.

[0134] In this embodiment, multiple texture-reconstructed images are obtained from the previous steps. These images should contain different road objects, such as vehicles, pedestrians, traffic signs, etc., to ensure the diversity and representativeness of the data. A suitable computer vision model is selected for road object recognition. A convolutional neural network (CNN) such as FasterR-CNN, YOLO or SSD is used. These models can quickly and accurately identify objects in the image. If there is no ready-made model, the model is trained using a training set containing labeled data to ensure that it has a good recognition effect. Each texture-reconstructed image is input into the trained object recognition model, and the inference process is performed. The model The model will analyze the image and mark all identified road object nodes, including the category and location coordinates of the object, record the location information of each object node, and generate corresponding annotation results for use in subsequent steps. Feature extraction is performed on each marked road object node, and the image area of ​​the object node is input into the selected deep learning model. Through forward propagation, the semantic feature vector of each object is extracted. These feature vectors include shape, color, texture and other information. The extracted road semantic feature vectors are recorded in the database to provide support for subsequent priority analysis and define the standards for road content priority, usually considering the type, size, location and impact of the object on traffic. For example, pedestrians and traffic lights typically have higher priority than other objects. Based on the road semantic features of each object node, a defined priority calculation method is applied to assign a priority score to each object node. A weighted algorithm is used to consider the influence of relevant features and select an appropriate edge detection algorithm, such as Canny edge detection, Sobel operator, or Laplacian operator. These algorithms can clearly extract the edge features of objects. The edge detection algorithm is executed on each marked road object node area to extract the edge detail features of the object and generate an edge image containing the outline and detail information of each object. An appropriate adaptive contrast enhancement algorithm, such as adaptive histogram equalization (CLAHE) or gamma adjustment, is selected. These algorithms can effectively improve the contrast of important areas in the image. For each object's edge detail feature map, the contrast enhancement parameters are dynamically adjusted based on the object's road content priority. For example, high-priority objects require stronger contrast enhancement, while low-priority objects receive moderate enhancement. Contrast enhancement is performed to generate multiple detail-enhanced images to ensure that the details of important objects are more prominent. The enhanced images are saved, and the parameters used in the enhancement process and the effect evaluation are recorded. These enhanced images should show significant improvements in detail representation to provide support for subsequent analysis and application.

[0135] In this embodiment, the specific steps of step S5 are:

[0136] Step S51: monitoring the real-time vibration parameters of the vehicle based on the vehicle-mounted sensors;

[0137] Step S52: performing time-series vibration feature evolution on the real-time vibration parameters of the vehicle to obtain the vehicle time-series vibration features;

[0138] Step S53: performing geometric posture vibration distortion analysis on the plurality of detail contrast enhanced images according to the vehicle time-series vibration characteristics, thereby generating a driving image vibration distortion characteristic;

[0139] Step S54: performing dynamic instantaneous distortion correction processing based on the vibration distortion characteristics of the driving image, thereby generating a dynamic distortion corrected image.

[0140] In this embodiment, suitable vehicle-mounted sensors, such as acceleration sensors, gyroscopes, or vibration sensors, are selected. These sensors can monitor the vibration data generated by the vehicle during driving in real time. The sensors are installed at appropriate locations on the vehicle, such as the chassis, wheels, or body, to ensure that the sensors can accurately capture the vibration information of the vehicle and avoid external interference. A real-time data acquisition system is developed to transmit the vibration parameters (such as acceleration, angular velocity, etc.) collected by the sensors to a central processing unit. Wireless communication technology (such as Bluetooth or Wi-Fi) is used for data transmission to ensure the real-time and accuracy of the data. The recorded vibration parameters include information such as timestamp, vibration amplitude, and frequency, forming a time series data set to provide a basis for subsequent analysis. The collected vibration data is preprocessed, including denoising and smoothing, to eliminate the influence of sensor noise. A filtering algorithm (such as a low-pass filter) is used to process the data to ensure data clarity. A time series analysis method, such as time series Fourier transform (FFT) or wavelet transform, is used to extract the frequency domain features and time features of the vibration parameters. These methods can reveal the regularity and change trend of the vibration data. The extracted time series vibration features, including frequency, The amplitude and phase information are recorded to form the vehicle's time-series vibration feature set. These features will be used for subsequent distortion analysis. The geometric distortion analysis method, such as affine transformation, perspective transformation or principal component analysis (PCA), is selected to evaluate the distortion caused by vibration in the image. The extracted time-series vibration features are combined with the image data to analyze the geometric distortion caused by vehicle vibration in the image. By mapping the vibration features to the image features, the distorted area caused by vibration is identified, and the vibration distortion features of the driving image are generated. The type, degree and impact range of the distortion in each image are recorded to provide information for subsequent dynamic correction. Based on this, select a suitable dynamic correction algorithm, such as an interpolation-based image reconstruction method, an optical flow method, or a real-time estimation method. These algorithms can perform real-time correction of images according to the vibration distortion characteristics. The vibration distortion characteristics of each detail contrast-enhanced image are used as input, and the selected correction algorithm is applied to perform dynamic instantaneous distortion correction. The correction parameters are dynamically adjusted according to the vibration characteristics to adapt to different vibration conditions. The corrected dynamic distortion correction images are saved, and the parameters used in the correction process and the effect evaluation are recorded. These corrected images should have improved visual quality to provide support for subsequent analysis and application.

[0141] In this embodiment, the specific steps of step S6 are:

[0142] Step S61: segmenting the dynamic distortion corrected image into continuous frames, thereby obtaining continuous frame corrected images at multiple angles;

[0143] Step S62: performing multi-angle image feature point recognition on the continuous frame corrected images at multiple angles, and extracting the image continuous frame feature points at each angle;

[0144] Step S63: performing geometric transformation registration on the feature points of the continuous frames of the images at each angle, thereby generating multi-frame registration data of the feature points;

[0145] Step S64: performing dynamic panoramic image stitching on the dynamic distortion corrected image based on the multi-frame registration data of the feature points, thereby generating a dynamic panoramic stitching image;

[0146] Step S65: performing local smoothing processing on the spliced ​​portions of the dynamic panoramic spliced ​​image, thereby constructing a panoramic behavior optimized image model.

[0147] In this embodiment, the corrected dynamic distortion correction images are obtained from the previous step. These images should include continuous frames taken from different angles of the vehicle to ensure data diversity and continuity. A suitable image segmentation algorithm is selected, such as an optical flow method or a frame difference method. These methods can effectively identify and extract continuous image frames. Continuous frames are extracted from the dynamic distortion correction images. The selected segmentation algorithm is used to segment the image sequence into continuous frame correction images at multiple angles, ensuring that the time interval between each frame is uniform for subsequent feature point recognition. The extracted continuous frame correction images at multiple angles are stored in a database for subsequent analysis and processing. A suitable feature point recognition algorithm is selected. , such as Harris corner detection, SIFT (Scale Invariant Feature Transform) or ORB (Oriented FAST and Rotated BRIEF), these algorithms can effectively identify key point features in the image, perform feature point recognition algorithm on the continuous frame correction image of each angle, extract the feature points in each image frame, and calculate its descriptor for subsequent registration processing, store the extracted feature points and their descriptors in the database, and provide support for subsequent geometric transformation registration. Select appropriate geometric transformation registration methods, such as RANSAC (Random Sampling Consensus Algorithm) or least squares method, which can effectively handle the differences between feature points. Matching, improve the robustness of registration, match the feature points of continuous frames of images at each angle, use the selected registration algorithm to determine the geometric transformation relationship between different frames, generate multi-frame registration data of feature points, record the relative position information of each frame, store the generated multi-frame registration data of feature points for subsequent dynamic panoramic image stitching, select a suitable stitching algorithm, such as the stitching method based on image registration or the multi-view geometric stitching method, which can effectively handle image stitching under different perspectives, use the multi-frame registration data of feature points to stitch the dynamic distortion corrected images, calculate the position of each frame of image in the panorama according to the registration data, perform image fusion, and generate dynamic Panoramic stitching images, save the generated dynamic panoramic stitching images, and record the parameters and effect evaluation used in the stitching process, select appropriate smoothing algorithms, such as Gaussian blur, bilateral filtering or local contrast adjustment, which can effectively eliminate the seams and visual discontinuities at the stitching points, perform local smoothing on the stitching points of the dynamic panoramic stitching images, dynamically adjust the smoothing parameters according to the characteristics of the seams to achieve the best effect, ensure a smooth transition of the stitching area, save the smoothed panoramic behavior optimization image model, and record the parameters and effect evaluation during the processing. These optimized images should have significantly improved visual quality to provide support for subsequent analysis and application.

[0148] In this embodiment, an image enhancement system for a vehicle-mounted image sensor is provided, which is configured to execute the above-described image enhancement method for the vehicle-mounted image sensor, including:

[0149] The adaptive exposure module is used to obtain driving recording images of the vehicle at multiple angles; the driving recording images of the vehicle at multiple angles are subjected to dynamic scene adaptive exposure adjustment, thereby obtaining multiple adaptive exposure driving images;

[0150] The filtering and denoising module is used to perform deep network noise recognition on multiple adaptive exposure driving images and perform dynamic filtering and denoising on each sub-image, thereby generating multiple filtered and denoised optimized images;

[0151] A texture reconstruction module is used to perform image texture analysis on multiple exposure sub-images and perform texture distortion reconstruction to construct multiple texture reconstructed images;

[0152] A contrast enhancement module is used to perform visual recognition of road objects on multiple texture-reconstructed images and perform adaptive contrast enhancement processing to construct multiple detail contrast-enhanced images;

[0153] The distortion correction module is used to monitor the real-time vibration parameters of the vehicle based on on-board sensors. Based on the real-time vibration parameters of the vehicle, the module performs geometric posture vibration distortion analysis and dynamic instantaneous distortion correction on multiple detail contrast-enhanced images to generate dynamic distortion-corrected images.

[0154] The panoramic stitching module is used to perform geometric transformation and registration on the dynamic distortion correction images and perform dynamic panoramic image stitching to construct a panoramic behavior optimization image model.

[0155] The present invention ensures the acquisition of high-quality driving recording images under different lighting conditions through dynamic scene adaptive exposure adjustment. The adaptive exposure driving images at multiple angles help to improve the brightness and contrast of the image, enhance the details and information content of the image, and combine deep network noise recognition with dynamic filtering noise reduction to effectively remove noise in the image, improve the clarity and quality of the image. The generated filtered noise reduction optimized image reduces noise and artifacts in the image, improves the visual effect and recognition accuracy of the image, and improves the details and texture of the image through texture distortion reconstruction, making the image more realistic and clear. The constructed texture reconstruction image reduces blur and distortion in the image, enhances the texture characteristics of the image, and improves the recognition and analysis capabilities of the image. The adaptive contrast enhancement processing is effective. It helps to highlight the details and edges in the image, enhance the contrast and visual effects of the image. The generated detail contrast enhanced image improves the readability and recognition of the image, making road objects more clearly visible. It performs distortion correction processing based on real-time vibration parameters to effectively eliminate the distortion in the image caused by vehicle vibration and improve the accuracy and stability of the image. The dynamic distortion correction image can provide a more realistic and clear perspective, providing a more reliable data basis for subsequent analysis and processing. The dynamic distortion correction image is subjected to geometric transformation alignment and panoramic stitching to generate a panoramic behavior optimization image model, providing a wider field of view and more complete information. The constructed panoramic behavior optimization image model helps the vehicle system better understand the surrounding environment and improve driving safety and efficiency.

[0156] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0157] The foregoing description is intended only to provide specific embodiments of the present invention, which are intended to enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest manner consistent with the principles and novel features disclosed herein.

Claims

1. An image enhancement method for a vehicle-mounted image sensor, characterized in that: The following steps are involved: Step S1: Acquire driving recorded images of the vehicle at multiple angles; perform dynamic scene adaptive exposure adjustment on the driving recorded images of the vehicle at multiple angles, thereby obtaining multiple adaptive exposure driving images; Step S2: performing deep network noise recognition on multiple adaptive exposure driving images, and performing dynamic filtering and noise reduction on each sub-image, thereby generating multiple filtered and noise-reduced optimized images; Step S3: performing image texture analysis on the multiple exposure sub-images and performing texture distortion reconstruction to construct multiple texture reconstructed images; Step S4: performing road object visual recognition on the multiple texture reconstructed images and performing adaptive contrast enhancement processing to construct multiple detail contrast enhanced images; Step S5: monitoring the real-time vibration parameters of the vehicle based on the on-board sensors; performing geometric posture vibration distortion analysis and dynamic instantaneous distortion correction processing on the multiple detail contrast enhanced images based on the real-time vibration parameters of the vehicle, thereby generating a dynamic distortion corrected image; Step S6: performing geometric transformation registration on the dynamic distortion correction image and performing dynamic panoramic image stitching to construct a panoramic behavior optimization image model; Among them, the specific steps of step S1 are: Step S11: Acquire driving record images of the vehicle at multiple angles; Step S12: performing scene depth semantic segmentation on each of the vehicle's driving record images from multiple angles to obtain scene semantic features of each image; Step S13: performing image region segmentation according to the scene semantic features of each image, thereby generating a plurality of different scene semantic regions; Step S14: analyzing the dynamic changes of ambient light in a plurality of different scene semantic areas to generate dynamic features of ambient light in different scenes; Step S15: performing dynamic scene adaptive exposure adjustment on multiple semantic areas of different scenes according to the dynamic characteristics of ambient light of different scenes, thereby obtaining multiple adaptive exposure driving images; The dynamic scene adaptive exposure adjustment is specifically as follows: Perform scene region classification on multiple different scene semantic regions to obtain sky regions, road regions, vehicle regions, and building regions; The dynamic characteristics of ambient light in different scenes specifically include: characteristics of ambient light brightness changes in the sky, characteristics of brightness distribution in the road area, characteristics of vehicle surface light, and local brightness intensity of buildings; Based on the brightness variation characteristics of the sky environment, the overexposure reduction processing is performed on the sky area to obtain an overexposure optimized image of the sky area; Identify the local brightness distribution of the road area to extract images of dark areas of the road; Performing high-exposure processing on the dark road area image according to the brightness distribution characteristics of the road area to generate a high-exposure road area image; Identify surface reflective points in the vehicle area to generate reflective feature points on the vehicle surface; Performing local exposure adjustment on the reflective feature points on the vehicle surface based on the light features of the vehicle surface to obtain a local exposure vehicle area image; Adaptively performing exposure compensation on the building area based on the local brightness intensity of the building to obtain an adaptively exposed building area image; The specific steps of step S4 are: Step S41: performing road object visual recognition on the multiple texture reconstructed images and marking multiple road object nodes; Step S42: mining road semantic features of multiple road object nodes to obtain road semantic features of each object node; Step S43: performing a road content priority analysis based on the road semantic features of each object node to generate a road content priority for each object node; Step S44: performing object edge detail mining on multiple road object nodes to generate edge detail features of each object; Step S45: performing adaptive contrast enhancement processing on the edge detail features of each object based on the road content priority of each object node, thereby constructing multiple detail contrast enhanced images; The specific steps of step S5 are: Step S51: monitoring the real-time vibration parameters of the vehicle based on the vehicle-mounted sensors; Step S52: performing time-series vibration feature evolution on the real-time vibration parameters of the vehicle to obtain the vehicle time-series vibration features; Step S53: performing geometric posture vibration distortion analysis on the plurality of detail contrast enhanced images according to the vehicle time-series vibration characteristics, thereby generating a driving image vibration distortion characteristic; Step S54: performing dynamic instantaneous distortion correction processing based on the vibration distortion characteristics of the driving image, thereby generating a dynamic distortion corrected image.

2. The image enhancement method of the vehicle-mounted image sensor according to claim 1, characterized in that: The specific steps of step S2 are: Step S21: performing deep network noise recognition on multiple adaptive exposure driving images to extract all noise points in the images; Step S22: performing noise distribution analysis on all noise points in the image to generate image noise distribution features; Step S23: dividing the image based on the image noise distribution characteristics to obtain multiple exposure sub-images; Step S24: calculating the noise distribution level of the multiple exposed sub-images, thereby obtaining the noise distribution level of each sub-image; Step S25: performing dynamic filtering and noise reduction on each sub-image based on the noise distribution level of each sub-image, thereby generating a plurality of filtered and noise-reduced optimized images.

3. The image enhancement method of a vehicle-mounted image sensor according to claim 1, characterized in that: The specific steps of step S3 are: Step S31: performing image texture analysis on multiple exposure sub-images to obtain image texture features of different noise backgrounds; Step S32: performing multi-scale convolution on the image texture features of different noise backgrounds to generate multi-scale texture convolution features; Step S33: performing detailed texture distortion identification on the plurality of filtered noise reduction optimized images, and marking image areas with texture distortion; Step S34: reconstructing the texture-distorted image region according to the multi-scale texture convolution feature to construct a plurality of texture-reconstructed images.

4. The image enhancement method of a vehicle-mounted image sensor according to claim 1, characterized in that: The specific steps of step S6 are: Step S61: segmenting the dynamic distortion corrected image into continuous frames, thereby obtaining continuous frame corrected images at multiple angles; Step S62: performing multi-angle image feature point recognition on the continuous frame corrected images at multiple angles, and extracting the image continuous frame feature points at each angle; Step S63: performing geometric transformation registration on the feature points of the continuous frames of the images at each angle, thereby generating multi-frame registration data of the feature points; Step S64: performing dynamic panoramic image stitching on the dynamic distortion corrected image based on the multi-frame registration data of the feature points, thereby generating a dynamic panoramic stitching image; Step S65: performing local smoothing processing on the spliced ​​portions of the dynamic panoramic spliced ​​image, thereby constructing a panoramic behavior optimized image model.

5. An image enhancement system for a vehicle-mounted image sensor, characterized in that: The method for performing the image enhancement method of the vehicle-mounted image sensor according to claim 1 comprises: The adaptive exposure module is used to obtain driving recording images of the vehicle at multiple angles; the driving recording images of the vehicle at multiple angles are subjected to dynamic scene adaptive exposure adjustment, thereby obtaining multiple adaptive exposure driving images; The filtering and denoising module is used to perform deep network noise recognition on multiple adaptive exposure driving images and perform dynamic filtering and denoising on each sub-image, thereby generating multiple filtered and denoised optimized images; A texture reconstruction module is used to perform image texture analysis on multiple exposure sub-images and perform texture distortion reconstruction to construct multiple texture reconstructed images; A contrast enhancement module is used to perform visual recognition of road objects on multiple texture-reconstructed images and perform adaptive contrast enhancement processing to construct multiple detail contrast-enhanced images; The distortion correction module is used to monitor the real-time vibration parameters of the vehicle based on on-board sensors. Based on the real-time vibration parameters of the vehicle, the module performs geometric posture vibration distortion analysis and dynamic instantaneous distortion correction on multiple detail contrast-enhanced images to generate dynamic distortion-corrected images. The panoramic stitching module is used to perform geometric transformation and registration on the dynamic distortion correction images and perform dynamic panoramic image stitching to construct a panoramic behavior optimization image model.

Images