A real-time image quality enhancement method for EMC test of vehicle-mounted camera

Abstract

The present application relates to the technical field of digital image processing, in particular to a real-time image quality enhancement method for EMC test of vehicle-mounted camera, comprising: acquiring real-time image data and historical image data; performing multi-branch collaborative decoupling processing on the real-time image data to generate a frequency domain purification graph, a core structure graph, a spatial domain purification graph and a time domain stability graph; performing local statistical analysis on the real-time image data to generate an artifact intensity graph and a perceptual saliency graph; under a multi-resolution Laplacian pyramid framework, fusing the frequency domain purification graph and the spatial domain purification graph; the weight of the fusion process is intelligently controlled in space and level by the artifact intensity graph, the perceptual saliency graph and the time domain stability graph based on a Gaussian pyramid, and finally a high-quality enhanced image is reconstructed. Through the multi-branch collaborative decoupling and intelligent fusion method, the real-time enhancement of the image of the vehicle-mounted camera under EMC interference is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital image processing, and particularly to a real-time image quality enhancement method for EMC test of a vehicle-mounted camera. BACKGROUND

[0002] The vehicle-mounted camera is a key visual sensor of the advanced driver assistance system (ADAS) and the autonomous vehicle, and its reliability is directly related to the driving safety. The electromagnetic compatibility (EMC) test is a necessary link to ensure the stable operation of the camera in a complex electromagnetic environment. In the test, the strong electromagnetic interference applied will seriously pollute the image signal, resulting in the degradation of the output image quality, which brings a unique challenge to the image processing field.

[0003] The existing image restoration methods are usually targeted at specific noise models. For example, the enhancement methods based on spatial domain filtering, such as Gaussian filtering or median filtering, have a certain smoothing effect on Gaussian noise or salt and pepper noise, but they are not good at filtering out periodic stripe noise or net noise, and will cause the blurring of high-frequency details such as edges and textures. On the other hand, the enhancement methods based on non-spatial domain filtering, such as the notch filter designed in the frequency domain by using Fourier transform, can effectively remove periodic noise of a specific frequency, but it is powerless to random distribution of wideband noise and time-domain noise, and is easy to introduce the ringing effect.

[0004] The prior art lacks a unified framework capable of processing mixed noise cooperatively, and cannot provide a comprehensive solution capable of responding to mixed and time-varying strong noise in the EMC test in real time and dynamically, nor can it protect the key visual structure by means of fine image analysis while effectively suppressing noise, which also leads to the difficulty and inaccuracy of test evaluation.

[0005] Therefore, the present application provides a real-time image quality enhancement method for EMC test of a vehicle-mounted camera to solve the above technical problems. SUMMARY

[0006] The present application aims to provide a real-time image quality enhancement method for EMC test of a vehicle-mounted camera to solve the problems raised in the background.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solution: a real-time image quality enhancement method for EMC test of a vehicle-mounted camera, comprising:

[0008] obtaining real-time image data and historical image data of the vehicle-mounted camera;

[0009] performing multi-branch collaborative decoupling on the real-time image data; a frequency domain decoupling branch, filtering EMC noise in the real-time image data in a frequency domain to generate a frequency domain purification image; a structure feature branch, extracting core structures from the frequency domain purification image to generate a core structure image; a spatial domain decoupling branch, performing smoothing processing on random noise in the real-time image data under the guidance of the core structure image to generate a spatial domain purification image; and a time domain decoupling branch, combining historical image data and the real-time image data to form time series data, analyzing high-energy pixels, and generating a time domain stable image;

[0010] performing local image statistical analysis and multi-feature extraction on the real-time image data to generate an artifact intensity map and a perceptual saliency map;

[0011] based on the frequency domain purification image, the spatial domain purification image, the time domain stable image, and the core structure image, performing multi-resolution information fusion by using a pyramid collaborative fusion algorithm, and adjusting and controlling fusion weights by using the artifact intensity map and the perceptual saliency map to generate an enhanced image.

[0012] Preferably, the frequency domain decoupling branch specifically comprises:

[0013] performing two-dimensional Fourier transform on the real-time image data to obtain a frequency spectrum image, performing peak value detection and conjugate symmetry analysis on the frequency spectrum image, setting a dynamic spectrum threshold according to an energy mean value and a standard deviation of the frequency spectrum image, identifying candidate noise frequency points with energy values exceeding the dynamic threshold and having conjugate symmetric distribution characteristics, calculating an energy concentration degree of each candidate noise frequency point, dynamically adjusting a center frequency and a filter bandwidth of a band-stop filter, filtering the candidate noise frequency points by using the band-stop filter, and performing two-dimensional inverse Fourier transform on all filtered candidate noise frequency points to generate the frequency domain purification image.

[0014] Preferably, the structure feature branch specifically comprises:

[0015] performing multi-scale gradient calculation on the frequency domain purification image to generate gradient amplitude maps and gradient direction maps at different scales, performing non-maximum suppression on the gradient amplitude maps to preliminarily locate edge candidate pixels, connecting edge candidate pixels with similar gradient directions to generate continuous edge segments, removing continuous edge segments with lengths less than a preset length threshold and curvatures greater than a preset curvature threshold, combining and screening the remaining continuous edge segments to generate the core structure image.

[0016] Preferably, the spatial domain decoupling branch specifically comprises:

[0017] Based on the core structure diagram, a structure region mask is generated through a morphological dilation operation; the structure region mask is inverted to generate a smooth region mask; based on the structure region mask, bilateral filtering processing is performed on the real-time image data to generate a first filtering result; based on the smooth region mask, Gaussian filtering processing is performed on the real-time image data to generate a second filtering result; the first filtering result and the second filtering result are merged to generate a spatial purification map.

[0018] Preferably, the time domain decoupling branch is specifically:

[0019] The real-time image data is combined with the historical image data to form time series data; motion estimation based on block matching is performed on the time series data to determine a corresponding matching block in the historical image data for each pixel block in the real-time image data, and a motion vector field is generated; based on the motion vector field, motion compensation is performed on the historical image data to generate compensated historical image data aligned with the real-time image data; based on the pixel brightness difference between the real-time image data and the compensated historical image data, a residual sequence is generated; one-dimensional Fourier transform is performed on the residual sequence to obtain a residual frequency spectrum; high-energy pixel coordinates are extracted from the residual frequency spectrum, and the high-energy pixel coordinates and the real-time image data are integrated to generate a time domain stability map.

[0020] Preferably, the generation process of the artifact intensity map and the perceptual saliency map is specifically:

[0021] The real-time image data is scanned in a local neighborhood, and the pixel value variance and gradient direction histogram entropy of the local neighborhood of the pixel block are calculated; the pixel value variance and gradient direction histogram entropy are weighted and summed to generate the artifact intensity map.

[0022] Parallel feature analysis is performed on the real-time image data, including linear feature analysis using Hough transform, color feature analysis using color range threshold segmentation, and shape feature analysis using contour detection, to generate a linear feature map, a color feature map, and a shape feature map; the linear feature map, the color feature map, and the shape feature map are weighted and fused, and the fusion weight is adjusted using the artifact intensity map to generate a perceptual saliency map.

[0023] Preferably, based on the frequency domain purification map, the spatial purification map, the time domain stability map, and the core structure diagram, multi-resolution information fusion is performed using a pyramid collaborative fusion algorithm, and the fusion weight is regulated through the perceptual saliency map and the artifact intensity map to generate an enhanced image, and the specific implementation process is:

[0024] The frequency domain purification map and the spatial purification map are respectively constructed into their corresponding Laplacian pyramids; the perceptual saliency map, the artifact intensity map, and the time domain stability map are respectively constructed into their corresponding Gaussian pyramids.

[0025] Based on the artifact intensity map and the time domain stability map in each level of the Gaussian pyramid, initial fusion weights are calculated for pixel positions in the corresponding level; the initial fusion weights are spatially modulated using the perceptual saliency map in the corresponding level of the Gaussian pyramid to enhance the weight of the image source that better preserves details in the perceptual salient region, thereby generating a final fusion weight map;

[0026] Using the final fusion weight map, the frequency domain purification map and the spatial domain purification map in the corresponding level of the Laplacian pyramid are weighted and fused to generate a fused Laplacian pyramid level; all the fused Laplacian pyramid levels are combined to generate a preliminary enhanced image;

[0027] Using the core structure map, edge enhancement is performed on the preliminary enhanced image to generate a final enhanced image.

[0028] Compared with the prior art, the present application has the following beneficial effects:

[0029] 1. The present application designs a brand-new system architecture, which can cooperatively and parallelly decompose and process complex image degradation problems in four dimensions of frequency domain, spatial domain, time domain and structure domain, and then cooperatively reconstruct all processing and analysis results through a multi-dimensional fusion algorithm. Compared with the single-dimensional island solution of the prior art, the system architecture of the present application fundamentally solves the inherent contradiction between artifact suppression and information fidelity. It can specifically process various artifacts without damaging the clarity of another information due to the suppression of one artifact, thereby achieving accurate suppression of composite and dynamic artifacts and maximizing the preservation of image information.

[0030] 2. The present application designs a serially dependent attention map generation process, which first generates an artifact intensity map, and then uses it as a kind of noise confidence signal to dynamically regulate the generation process of the perceptual saliency map. It can intelligently distrust the features detected in the noise serious area, thereby effectively avoiding the phantom salient target caused by noise. This ensures that the control signal used for guidance of fusion is accurate and directly improves the intelligence of the whole system.

[0031] 3. In the final fusion stage of the present application, a multi-step fusion process including two-stage weight generation and structure post-compensation using the pyramid cooperative fusion algorithm is adopted. The two-stage weight generation using the pyramid cooperative fusion algorithm makes the fusion decision consider both physical properties (noise, stability) and semantic importance (ADAS saliency), which is highly intelligent and context-aware. The structure post-compensation step uses the reliable core structure map for final sharpening, thereby ensuring that the edges and details of the output image reach the maximum fidelity in physics. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A flow chart of a real-time image quality enhancement method for vehicle-mounted camera EMC testing is provided for the embodiments of the application;

[0033] Figure 2 A structural schematic diagram of a multi-branch cooperative decoupling architecture is provided for the embodiments of the application;

[0034] Figure 3 A work flow chart of a pyramid cooperative fusion algorithm is provided for the embodiments of the application. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, but not all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work are within the protection scope of the application.

[0036] The application provides a real-time image quality enhancement method for vehicle-mounted camera EMC testing, and the technical solutions are as follows:

[0037] Real-time image data and historical image data of the vehicle-mounted camera are acquired;

[0038] The real-time image data is subjected to multi-branch cooperative decoupling; a frequency domain decoupling branch is used to filter EMC noise in the real-time image data in the frequency domain to generate a frequency domain purification image; a structural feature branch is used to extract core structures from the frequency domain purification image to generate a core structure image; a spatial domain decoupling branch is used to smooth random noise of the real-time image data under the guidance of the core structure image to generate a spatial domain purification image; and a time domain decoupling branch is used to combine the historical image data and the real-time image data to form time series data, analyze high-energy pixels, and generate a time domain stable image;

[0039] Local image statistical analysis and multi-feature extraction are performed on the real-time image data to generate an artifact intensity image and a perceptual saliency image;

[0040] Based on the frequency domain purification image, the spatial domain purification image, the time domain stable image, and the core structure image, multi-resolution information fusion is performed by using a pyramid cooperative fusion algorithm, and the fusion weight is regulated and controlled by using the artifact intensity image and the perceptual saliency image to generate an enhanced image.

[0041] The technical solutions of the application will be further described in detail below with reference to specific embodiments.

[0042] Embodiment One

[0043] This embodiment takes a typical vehicle-mounted camera EMC compatibility test scene as an embodiment background.

[0044] Referring Figure 1 The present application proposes specific implementation steps of the method, which include:

[0045] S1: obtaining real-time image data and historical image data of the vehicle-mounted camera;

[0046] S2: performing multi-branch collaborative decoupling on the real-time image data; a frequency domain decoupling branch filters out EMC noise in the real-time image data in the frequency domain to generate a frequency domain purification image; a structural feature branch extracts core structures from the frequency domain purification image to generate a core structure image; a spatial domain decoupling branch performs smoothing processing on random noise in the real-time image data under the guidance of the core structure image to generate a spatial domain purification image; and a time domain decoupling branch combines the historical image data and the real-time image data to form time series data, analyzes high-energy pixels, and generates a time domain stable image;

[0047] S3: performing local image statistical analysis and multi-feature extraction on the real-time image data to generate an artifact intensity map and a perceptual saliency map;

[0048] S4: based on the frequency domain purification image, the spatial domain purification image, the time domain stable image, and the core structure image, performing multi-resolution information fusion using a pyramid collaborative fusion algorithm, and adjusting and controlling the fusion weight through the artifact intensity map and the perceptual saliency map to generate an enhanced image.

[0049] Further, the real-time image data and the historical image data of the vehicle-mounted camera are obtained, corresponding to the above S1 step, and the specific process is as follows:

[0050] A high-speed image acquisition device is used to obtain the original video stream of the vehicle-mounted camera under high resolution and high frame rate as real-time image data. In this embodiment, a car front-view camera with a GMSL interface is used for image acquisition at a resolution of 1920x1080 pixels and a frame rate of 60 frames / second. Meanwhile, the video images of the previous 60 frames are cached as historical image data for subsequent time domain analysis. These data streams are transmitted to the FPGA integrated processor through the PCIe Gen3x4 interface.

[0051] Further, the real-time image data is subjected to multi-branch collaborative decoupling, including a frequency domain decoupling branch, a structural feature branch, a spatial domain decoupling branch, and a time domain decoupling branch, corresponding to the above S2 step, and the specific process is as follows: the four decoupling branches work collaboratively to process EMC artifacts in the image from different dimensions.

[0052] Specifically, the frequency domain decoupling branch filters out EMC noise in the real-time image data in the frequency domain to generate a frequency domain purification image, and the specific process is as follows:

[0053] The real-time image data collected is subjected to two-dimensional Fourier transform to convert the image from spatial domain to frequency domain, thereby obtaining a frequency spectrum. In the frequency spectrum, periodic noise is usually manifested as discrete, high-energy bright spots, and due to the characteristics of Fourier transform, these bright spots are symmetrically distributed with the origin.

[0054] The acquired frequency spectrum is analyzed to calculate the overall energy mean and standard deviation of the frequency spectrum. A dynamic spectrum threshold is set according to the overall energy mean and standard deviation. Any frequency point whose energy value exceeds the dynamic spectrum threshold and exhibits conjugate symmetric distribution characteristics is identified as a preliminary candidate noise frequency point. The historical image data is motion compensated, and the stability of the preliminary candidate noise frequency point is analyzed. The temporal stability is measured according to the number of frames in the historical image data that contain the corresponding preliminary candidate noise frequency point. In this embodiment, if there are 20 frames in the first 30 frames of historical image data that contain the corresponding preliminary candidate noise frequency point, it is considered that the preliminary candidate noise frequency point has stability and is further screened as a candidate noise frequency point.

[0055] The candidate noise frequency point is identified by using a combination of intra-frame double standard preliminary screening and motion compensation and temporal verification. This layer-by-layer verification mechanism ensures that the frequency point finally identified as noise has a very high confidence level, significantly reducing the false positive rate.

[0056] Each candidate noise frequency point and its immediate neighborhood are examined. The energy concentration degree is quantified by calculating the normalized kurtosis of the field to determine whether the energy is highly concentrated at the center point or more evenly distributed throughout the region. If the energy concentration degree is high, it indicates that the frequency point is indeed significant periodic noise. The energy concentration degree of each candidate noise frequency point is calculated. Specifically, taking the candidate noise frequency point as the center, a N×N neighborhood is taken (N is an odd number, preferably in the range of [3, 7], and in this embodiment, N = 3), the kurtosis of the energy distribution in the neighborhood is calculated, and the kurtosis value is taken as the quantization result of the energy concentration degree. The energy concentration degree and the time stationarity are normalized and then weighted and fused to generate a noise confidence.

[0057] BW = BW min +(BW max -BW min )×Sigmoid(C norm ),

[0058]

[0059] where BW min and BW max are the preset minimum and maximum bandwidths, respectively, and in this example, they are set to 1 frequency unit and 5 frequency units, respectively, and C norm is the normalized noise confidence.

[0060] The energy attenuation factor is set according to the noise confidence, and the center frequency and filter bandwidth of the band-stop filter are adjusted according to the energy concentration.

[0061] The present application fuses the spectrum peak shape and the time dimension for cross verification, generates the noise confidence to attenuate and adjust the filter. It can effectively filter out some transient artifacts. The spectrum peak must pass through the strict examination of the morphology and time sequence two checkpoints to be given a higher noise confidence. This significantly reduces the misjudgment rate of noise recognition, ensures that the subsequent filtering operation is only for the real and persistent EMI interference, and fundamentally improves the accuracy and reliability.

[0062] The band-stop filter is used to filter the candidate noise frequency points in the spectrum graph, that is, to greatly attenuate these high-energy noise components in the frequency domain. When there are multiple periodic interference frequencies, the filtered candidate noise frequency points are integrated, and then two-dimensional inverse Fourier transform is performed to convert the image data from the frequency domain to the spatial domain, thereby generating a frequency domain purification image that significantly removes periodic artifacts.

[0063] Compared with fixed frequency domain filtering, the frequency domain decoupling branch in the present method realizes intelligent and accurate identification of periodic noise through symmetry analysis and dynamic threshold, and realizes accurate filtering of noise and maximum protection of details through dynamic adjustment of bandwidth.

[0064] Specifically, the structural feature branch extracts the core structure from the frequency domain purification image to generate a core structure image, specifically:

[0065] Multi-scale gradient calculation is performed on the frequency domain purification image generated by the frequency domain decoupling branch. This means that not only the intensity and direction of the change in pixel brightness in a fixed size area are calculated, but also larger scale calculations are performed. The present embodiment uses a 3x3 size Sobel operator and a 5x5 size Sobel operator for multi-scale gradient calculation. In this way, edges of different thickness and intensity in the image can be more comprehensively captured, and gradient amplitude maps and gradient direction maps at different scales can be generated.

[0066] Non-maximum suppression (NMS) operation is performed on the gradient amplitude map. By comparing the gradient amplitude of the edge pixel point with the gradient amplitude of the adjacent pixels in its gradient direction, only the local maximum value point is retained, so that the blurred edge line is thinned to a single-pixel-width edge line, and the accurate edge candidate pixels are preliminarily located.

[0067] The refined edge candidates are traversed and connected to form continuous edge segments based on their spatial positions and gradient direction similarities. In this embodiment, two edge pixels are considered to belong to the same edge line if the distance between them is less than 5 pixels and the difference in their gradient directions is less than 15 degrees.

[0068] The generated continuous edge segments are strictly filtered. Those with length less than a pre-set length threshold (10 pixels in this embodiment) and curvature greater than a pre-set curvature threshold (0.5 radian / pixel in this embodiment) are removed. These removed segments are usually considered as noise residuals or unimportant details. The remaining continuous edge segments after filtering are combined to generate a binary core structure map. This map accurately depicts the key and visually salient structure information in the image, such as lane lines, vehicle outlines, road signs, etc.

[0069] Compared to directly extracting structures from a noisy image, the structure feature branch in this method accurately extracts and generates the core structure map of the image through multi-scale gradient calculation, non-maximum suppression, edge connection, and filtering based on length and curvature on the frequency domain cleaned map. This ensures that the visual clarity and structural integrity of the image are maximally preserved while being strongly denoised, preventing image distortion caused by over-denoising.

[0070] Specifically, the spatial decoupling branch smoothes the random noise in the real-time image data under the guidance of the core structure map to generate a spatially cleaned map, which is:

[0071] The core structure map generated by the structure feature branch. A structure region mask is generated by performing a morphological dilation operation on the core structure map. The mask covers all important edge and structure regions in the image. The structure region mask is inverted to generate a smoothing region mask, which covers the remaining non-structure regions in the image, mainly the background and flat regions.

[0072] According to the two masks, differential filtering is performed on the original real-time image data. In the structure region, bilateral filtering is performed on the real-time image data to generate the first filtering result. Because it considers not only the spatial distance between pixels but also the similarity of their pixel values when calculating the weighted average of pixels, bilateral filtering can effectively preserve the edges of the image while smoothing noise. When processing lane line edges, if the pixel value difference is large, even if the spatial distance is close, it will not be excessively smoothed.

[0073] In the smoothing region, Gaussian filtering is performed on the real-time image data. Gaussian filtering performs a weighted average of pixels, and the weights follow a Gaussian distribution, which has a significant advantage in denoising flat regions.

[0074] Finally, the first filtered result (structure region) and the second filtered result (smooth region) are seamlessly merged according to the respective masks to generate a spatially cleaned image that effectively removes random noise while accurately preserving the core structure details of the image.

[0075] Compared with a single global filter, the spatial decoupling branch in this method uses the most suitable filter for different regions of the image through a partitioned differential filtering strategy based on the structure mask, effectively balancing between strong noise removal (flat area) and edge preservation (structure area).

[0076] Specifically, the temporal decoupling branch combines historical image data and real-time image data to form time series data, analyzes high-energy pixels, and generates a temporal stability map, which is:

[0077] The real-time image data of the current frame is combined with the historical image data obtained from the cache to form a short-term time series data sequence.

[0078] Motion estimation based on block matching is performed on the time series data. The real-time image data is divided into small pixel blocks (16x16 pixel blocks are used in this embodiment), then Census transformation is performed to convert the brightness domain to a brightness-insensitive feature encoding domain, and for each pixel block in the real-time image data, the most similar feature matching block in the historical image data after Census transformation in the previous frame is searched to determine the motion vector of the pixel block. This process generates a motion vector field that accurately describes the motion of image content between frames. Using this motion vector field, the historical image data is motion compensated by moving the pixels in the historical image to the corresponding position of the real-time image according to their motion vectors, generating compensated historical image data that is accurately aligned with the real-time image data.

[0079] This solution decouples the two visual detection problems of motion and illumination by mapping the image in the brightness domain to features for motion matching. It ensures that even when the lighting environment changes dramatically, the motion vector can be calculated stably, providing a solid foundation for subsequent motion compensation.

[0080] The pixel intensity difference between the real-time image data and the motion compensated historical image data is calculated to generate a residual sequence. This residual sequence mainly reflects the image changes caused by non-motion, which includes the flicker and transient fluctuation caused by EMC interference. A one-dimensional Fourier transform is performed on the residual sequence to obtain its residual frequency spectrum. In this frequency spectrum, the flicker usually appears as a specific high-energy frequency component, for example, if the interference frequency is 50Hz, there will be a high-energy peak at the 50Hz position of the residual frequency spectrum. The high-energy pixel coordinates are identified and extracted according to the frequency spectrum. The energy mean value of each pixel's own residual frequency spectrum is calculated, and a residual dynamic judgment threshold is set, which is N times of its own energy mean value, where N is an adjustable gain coefficient, preferably in the range of [3, 5]. The residual frequency spectrum of the pixel is traversed, and if there is an energy peak in a specific interference frequency band (such as 50Hz-150Hz), and the energy of the peak exceeds the residual dynamic judgment threshold, the coordinates of the pixel are recorded as high-energy pixel coordinates.

[0081] According to the coordinates of these high-energy pixels, adjustments are made in combination with real-time image data. For continuous flicker, Kalman filtering is performed on these pixels in the time dimension to smooth periodic fluctuations. For transient high-energy pixels, compensation may be made through time domain interpolation and fast repair mechanisms. Finally, a time domain stable image is generated that effectively suppresses flicker and transient fluctuations and has stronger visual coherence.

[0082] Compared with simple time domain variance analysis, the time domain decoupling branch in this method combines real-time and historical image data to form time series data, performs motion estimation and motion compensation based on block matching, and analyzes the frequency spectrum of the residual sequence to extract high-energy pixels. By utilizing inter-frame time series information and motion compensation, flicker caused by low-frequency interference can be effectively suppressed, and image abnormalities caused by transient high-energy impact can be smoothed or compensated, generating a time domain stable image that enhances the coherence and smoothness of the video stream.

[0083] Further, local image statistical analysis and multi-feature extraction are performed on the real-time image data to generate an artifact intensity map and a perceptual saliency map, corresponding to the S3 step, the specific process is as follows:

[0084] The real-time image data is scanned in local neighborhood, in this example, a 5x5 pixel size sliding window is used to traverse every pixel block in the image. The variance of the intensity values of all pixels inside the pixel block is calculated. The larger the variance, the more fluctuation of the pixel values in this region, which usually means more random noise and artifacts. The histogram entropy of the gradient direction of the pixels inside the pixel block is calculated. The gradient direction histogram can reflect the complexity of the local texture and edges, if the entropy value is high, it means the gradient direction distribution of this region is chaotic, which is often caused by the increase of noise. The pixel value variance map and the gradient direction histogram entropy map are weighted and summed to generate the artifact intensity map. The sum of the weights of the pixel value variance map and the gradient direction histogram entropy map is 1, and the weight of the pixel value variance map is preferably in the range of [0.5, 0.7]. In this map, the higher the pixel value, the stronger the artifacts or noise in this region; the lower the pixel value, the better the image quality in this region.

[0085] The real-time image data is analyzed in parallel features. The Hough transform is used to detect straight lines in the image, and the detected straight line regions are assigned high pixel values in the linear feature map, indicating their importance as linear features. Based on the preset color range threshold segmentation, the identified specific color regions are assigned high pixel values in the color feature map, indicating their importance as color features. The contour detection algorithm is used to analyze the closed contours or complex shapes in the image, and the detected object contour regions are assigned high pixel values in the shape feature map, indicating their importance as shape features.

[0086] The generated linear feature map, color feature map and shape feature map are preliminarily weighted and fused. Then, the preliminary fused weight is spatially modulated using the previously generated artifact intensity map to generate the perceptual saliency map.

[0087] S final (p)=S init (p)×(1-A norm (p)),

[0088] where S final is the modulated result, S init is the original perceptual saliency map, p is the pixel position, and A norm is the artifact intensity value normalized in the interval [0, 1].

[0089] If the value of a region in the artifact intensity map is very high (indicating that the noise in this region is severe), the corresponding feature weight of this region in the fused weight is reduced. This avoids misjudging noise and artifacts as regions with high perceptual saliency, ensuring the accuracy of the final perceptual saliency map. The finally generated perceptual saliency map highlights the truly important, clear and high visual information part of the image in the highlight area.

[0090] Compared with generating attention map independently, the method generates artifact intensity map by local neighborhood scanning, pixel value variance and gradient entropy calculation, and generates perceptual saliency map by linear, color and shape feature analysis and fusion with artifact intensity map. The artifact intensity map can quantify the noise and artifact degree of the local region of the image, and the perceptual saliency map can highlight the most important region in the image for the human eye or subsequent automatic driving perception task, so as to ensure that the image quality in these key regions is given priority to protect and avoid important information being over-smoothed.

[0091] Further, based on the frequency domain purification map, the spatial domain purification map, the time domain stability map and the core structure map, a pyramid collaborative fusion algorithm is used for multi-resolution information fusion, and the fusion weight is regulated by the artifact intensity map and the perceptual saliency map to generate an enhanced image. Corresponding to the above S4 step, the specific process is as follows:

[0092] For the frequency domain purification map and the spatial domain purification map, a corresponding Laplacian pyramid is constructed. The Laplacian pyramid decomposes the image at different spatial frequencies and scales to obtain detailed information, and the high-level pyramid retains rough information and the low-level pyramid contains fine details. In this embodiment, a 5-layer pyramid is constructed to cover various scale information from coarse to fine.

[0093] For the perceptual saliency map, the artifact intensity map and the time domain stability map, a corresponding Gaussian pyramid is constructed. The Gaussian pyramid generates summarized information of the image at different scales through successive smoothing and downsampling. In this embodiment, a 5-layer Gaussian pyramid corresponding to the number of layers of the Laplacian pyramid is constructed to provide multi-scale fusion weights and stability information.

[0094] Based on the artifact intensity map and the time domain stability map in the current level Gaussian pyramid, the initial fusion weight W is calculated for each pixel position in the corresponding level.

[0095] W(p) = clip(a A(p) + b (1-T(p)), 0, 1),

[0096] Where a and b are hyperparameters that control the influence degree of artifact intensity and time domain instability, respectively, a+b=1, a is preferably in the range of [0.4, 0.6], and in this embodiment, a is set to a=0.5. The clip function truncates the results greater than 1 and less than 0, and limits the results to the range of [0, 1].

[0097] The initial fusion weight is spatially modulated by the perceptual saliency map in the corresponding level of the Gaussian pyramid. The fusion weight is further adjusted in the perceptual salient region, giving priority to the image source that can better preserve the details and structural information. This ensures that in the visually important region, the image details will not be excessively smoothed, thereby generating the final fusion weight map. Specifically, according to the pixel value in the perceptual saliency map, each pixel in the image is divided into a salient region and a non-salient region. For each pixel in the salient region, it is necessary to determine which of the two source images, "frequency domain purification map" and "spatial domain purification map", better preserves the details around the pixel. The evaluation method is to calculate the gradient amplitude of the pixel in a small neighborhood (such as a 3x3 or 5x5 pixel window) in the two source images respectively. The greater the gradient amplitude, the richer and clearer the edge or texture details of the region. Calculate the difference between the gradient amplitudes of the two maps, and map the difference to the [0, 1] interval using the Sigmoid function to form the detail preservation weight. For pixels in the non-salient region, the initial fusion weight is the final fusion weight; for pixels in the salient region, the final fusion weight is weighted average of the initial fusion weight and the detail preservation weight, and the weight is based on the pixel perceptual saliency value. This regulation mechanism makes the fusion result both effectively denoising and highlighting key information.

[0098] The final fusion weight map is used to weight and fuse the frequency domain purification map and the spatial domain purification map in the corresponding level of the Laplacian pyramid. For each pixel at each level, the fused Laplacian pyramid value is the weighted sum of the corresponding Laplacian pyramid values of the frequency domain purification map and the spatial domain purification map, and the weight is provided by the final fusion weight map. This step organically combines the detail information of the two different denoising results to generate the fused Laplacian pyramid level.

[0099] By inverse Laplace transform, all the fused Laplacian pyramid levels are combined to generate a preliminary enhanced image. The preliminary enhanced image has fused the denoising advantages of each branch and has been optimized in terms of perception and artifact degree. Finally, the core structure map generated in step S2 is used to perform edge enhancement on the preliminary enhanced image to generate the final enhanced image. Through non-sharpening mask technology, the core structure map is used as a mask to perform local contrast enhancement on the edge region of the preliminary enhanced image.

[0100] Compared with simple image fusion, the method is based on multi-layer pyramid construction, uses artifact intensity map, time domain stability map and perceptual saliency map to regulate the fusion weight in hierarchy and space, and finally performs edge enhancement. Using the pyramid structure, the information is fused at different scales, so that the denoising effect is more comprehensive. Through the guidance of artifact intensity map and time domain stability map, it is ensured that in the high noise and unstable area, stronger denoising weight is given; through the spatial modulation of perceptual saliency map, it can be more inclined to retain details in important visual areas, and the best balance between denoising and detail retention is realized. Finally, the core structure map is used for edge enhancement, which further improves the visual clarity and recognition of the image, so that the enhanced image is more suitable for subsequent machine vision processing.

[0101] The present application realizes real-time, adaptive and multi-dimensional suppression and enhancement of image artifacts caused by complex electromagnetic interference of the vehicle-mounted camera in EMC test through multi-branch cooperative decoupling (frequency domain, structural features, spatial domain, time domain) and multi-resolution pyramid cooperative fusion. The problem that the existing general image processing method cannot effectively and specifically remove complex artifacts such as periodic stripes, random noise, flicker and transient distortion caused by EMC interference is solved. Through decoupling parallel processing and pyramid fusion, real-time processing of high-resolution and high-frame-rate video streams is realized, solving the problem of low efficiency of traditional methods, making real-time monitoring in EMC test possible. Through the regulation of artifact intensity map and perceptual saliency map on the fusion weight, the method can intelligently adapt to the local characteristics of the image content and the artifact intensity, improving the processing robustness under different EMC interference types and intensities.

[0102] Embodiment two

[0103] Reference Figure 2 In this embodiment, the vehicle-mounted wide-angle front-view camera in EMC test of electric vehicles is taken as an example, and the image data of the vehicle-mounted wide-angle front-view camera is decoupled from different dimensions through four decoupling branches working cooperatively.

[0104] In the frequency domain decoupling branch, the two-dimensional Fourier transform is performed on the current frame image data obtained in real time, so that the image is converted from the spatial domain to the frequency domain, and the frequency spectrum of the image is obtained. In the frequency spectrum, the periodic bright lines usually appear as discrete and high-energy bright spots at specific frequency positions, and these bright spots are symmetrically distributed around the spectrum center (zero frequency). In order to intelligently identify these noise frequency points, the system first calculates the overall energy mean and standard deviation of the current frequency spectrum, and then calculates the dynamic spectrum threshold. The calculation formula of the dynamic spectrum threshold T is:

[0105] T = μ E +k·σ E ,

[0106] where μ Eis the average energy of the spectrogram, σ E is the standard deviation of the energy, k is a preset gain coefficient. To balance the false positive rate and false negative rate, the value of k is preferably in the range of [2.5, 4.0]. The spectrogram is traversed, and any frequency point with energy exceeding the dynamic spectrum threshold and exhibiting conjugate symmetric distribution characteristics is identified as a candidate noise frequency point.

[0107] For each identified candidate noise frequency point, the system further calculates its energy concentration. The energy concentration is measured by the kurtosis value. The center frequency and the filter bandwidth of the band-stop filter are dynamically adjusted according to the energy concentration. For the detected periodic noise with a frequency of 120Hz, it corresponds to a specific frequency point in the spectrogram. The center frequency of the band-stop filter is precisely set to the coordinate of the noise frequency point. The filter bandwidth is the weighted sum of the reference bandwidth, the energy concentration, and the difference between the current energy concentration and the maximum energy concentration. The adjusted band-stop filter is used to accurately filter the candidate noise frequency points in the spectrogram, i.e., to significantly attenuate these high-energy noise components in the frequency domain. Since the DC-DC converter may cause harmonic interference, if multiple periodic interference frequencies (e.g., 120Hz and its harmonics 240Hz) are detected, all filtered candidate noise frequency points are integrated. Finally, a two-dimensional inverse Fourier transform is performed to convert the image data from the frequency domain back to the spatial domain, generating a frequency domain purification image that significantly removes periodic bright line artifacts.

[0108] A multi-scale gradient calculation is performed on the frequency domain purification image generated by the frequency domain decoupling branch. This embodiment uses a 3x3 Sobel operator and a 5x5 Sobel operator for multi-scale convolution operations to capture edges of different thicknesses in the image, generating gradient magnitude maps and gradient direction maps at different scales. For example, for elongated structures such as lane lines, the 3x3 operator may be more sensitive; for thicker structures such as vehicle outlines, the 5x5 operator may provide more stable gradients. A breadth-first search algorithm is used to traverse these refined edge candidate pixels, and they are connected according to their spatial positions and gradient direction similarities to form continuous edge segments. Specifically, starting from an unvisited edge pixel P0, it is added to the queue. Then, a pixel Pi is taken out of the queue, and all unvisited edge pixels Pj in the neighborhood are checked. If the Euclidean distance between pixels Pj and Pi is less than 5 pixels and the absolute value of the difference in their gradient directions is less than 15 degrees, then Pj is marked as visited and added to the current edge segment and the queue. This process is repeated until the queue is empty, forming a complete continuous edge segment. The generated continuous edge segments are strictly screened. The continuous edge segments remaining after screening are combined to generate a binary core structure map.

[0109] The core structure map is generated by the structural feature branch in the spatial decoupling branch. A morphological dilation operation is performed on the core structure map using a 3x3 rectangular structuring element to generate a structure region mask. This mask covers all the important edges and structure regions in the image and expands slightly outward to include the pixels near the edges. An inverse operation is performed on the structure region mask to generate a smooth region mask, which covers the rest of the non-structure regions in the image, mainly the background and flat regions.

[0110] According to the two masks, a differential filtering process is performed on the original real-time image data:

[0111] In the region covered by the structure region mask, a bilateral filtering process is performed on the real-time image data. Bilateral filtering can effectively preserve the edges of the image while smoothing the noise, because it considers not only the spatial distance between pixels but also the similarity of pixel values when calculating the weighted average of pixels. In this embodiment, the spatial domain standard deviation of the bilateral filter is set to 5 and the pixel value domain standard deviation is set to 20. In the region covered by the smooth region mask, a Gaussian filtering process is performed on the real-time image data. Gaussian filtering performs a weighted average of pixels, and the weights follow a Gaussian distribution, which has a significant advantage in denoising flat regions. In this embodiment, the standard deviation of the Gaussian filter is set to 2.

[0112] Finally, the bilateral filtering result of the structure region and the Gaussian filtering result of the smooth region are seamlessly merged according to their respective masks to generate a spatial purification map that effectively removes random noise and accurately preserves the details of the core structure of the image.

[0113] In the temporal decoupling branch, the real-time image data of the current frame is combined with the historical image data of the previous 60 frames obtained from the cache to form a short-term time series data sequence. Motion estimation based on block matching is performed on the time series data. The real-time image data is divided into small blocks of 16x16 pixels. For each pixel block in the real-time image, the most similar matching block is searched in the historical image data to determine the motion vector of the pixel block. The metric standard of block matching uses the mean absolute difference, which has relatively small computational complexity and is somewhat robust to noise. The search strategy uses a three-step search method to balance search speed and accuracy. This process generates a motion vector field that accurately describes the motion of image content between frames.

[0114] The motion vector field is used to perform motion compensation on the historical image data, i.e. the pixels in the historical image are shifted to the corresponding positions in the real-time image according to their motion vectors, to generate compensated historical image data that is accurately aligned with the real-time image data. The pixel intensity difference between the real-time image data and the motion compensated historical image data is calculated to generate a residual sequence. The residual sequence mainly reflects the image changes caused by non-motion, including the flicker and transient fluctuations caused by EMC interference. A one-dimensional Fourier transform is performed on the residual sequence to obtain its residual frequency spectrum, which has a frequency resolution of 1 Hz. The energy peak in a preset typical flicker frequency band (e.g. 40 Hz-130 Hz) is searched. If a peak point with energy exceeding a dynamic threshold set according to the mean and standard deviation of the pixel residual energy is detected, the coordinates of the pixel are recorded as the high-energy pixel coordinates, and the frequency at which the peak value is located is taken as the flicker frequency of the point.

[0115] According to the high-energy pixel coordinates, Kalman filtering is performed on the high-energy pixels to smooth the periodic fluctuations. The state variable of the Kalman filter is defined as the true intensity value of the pixel and its change rate. For transient high-energy pixels (e.g. sharp bright or dark spots that appear for a short time), compensation is performed through time-domain median filtering and a fast repair mechanism based on the neighborhood. Specifically, for a pixel point identified as a transient high-energy pixel, its value is replaced by the median of the corresponding position pixels in the adjacent two frames. Finally, a time-domain stable image is generated that effectively suppresses flicker and transient fluctuations and has stronger visual coherence.

[0116] Embodiment Three

[0117] This embodiment uses local image statistical analysis and multi-feature extraction to generate an artifact intensity map and a perceptual saliency map in a truck reversing assistance system.

[0118] The real-time image data (e.g. a scene image captured by the current frame containing the road conditions behind the truck, the stop line and a small amount of scattered objects, in which the image presents obvious random snowflake-like noise due to EMC interference) is scanned in local neighborhoods. In this embodiment, a 5x5 pixel size sliding window is used to traverse each pixel block of the image.

[0119] For each pixel block covered by the sliding window, the system calculates the variance of the intensity values of all the pixels inside it. The larger the variance, the greater the fluctuation of the pixel values in the region, which usually means that the random noise and artifacts are stronger. At the same time, the system also calculates the entropy of the gradient direction histogram of the pixels inside the pixel block. The gradient direction histogram can reflect the complexity of local texture and edges; if the entropy value is high, it means that the gradient direction distribution of the region is chaotic, which is often an increase in chaos caused by random noise.

[0120] To generate the artifact intensity map, the system will perform a weighted sum of the computed pixel value variance map and the gradient direction histogram entropy map. In this embodiment, the pixel value variance map is given a relatively higher weight, set to 0.6, while the gradient direction histogram entropy map is given a weight of 0.4. This means that areas with high pixel value fluctuation will contribute more to the artifact intensity map. In this artifact intensity map, areas with higher pixel values indicate that the region has higher artifact or noise; areas with lower pixel values indicate that the region has relatively better image quality. For example, in areas with severe snowflake noise, the pixel values of the artifact intensity map will be significantly higher.

[0121] Parallel feature analysis is performed on the real-time image data to identify regions in the image that are of high visual or semantic importance to human drivers or autonomous driving systems. These analyses include: linear feature analysis, which utilizes the Hough transform to detect straight lines in the image. Detected straight line regions will be given high pixel values in the linear feature map. color feature analysis, which utilizes color range thresholding for color feature analysis. In a reversing scenario, specific traffic sign colors (e.g. red stop signs, yellow warning lines), bright colors of obstacles, etc. can be important color features. In this embodiment, the color ranges of interest to the system are set to be red hues (e.g. high red channel luminance and relatively low green and blue channel luminance) and yellow hues (e.g. high red and green channel luminance and relatively low blue channel luminance). These identified specific color regions will be given high pixel values in the color feature map. shape feature analysis, which utilizes a contour detection algorithm to analyze closed contours or complex shapes in the image. Detected object contour regions will be given high pixel values in the shape feature map.

[0122] The system will perform a preliminary weighted fusion of the generated linear feature map, color feature map, and shape feature map to obtain a preliminary perceptual saliency map. In this embodiment, the linear feature map, color feature map, and shape feature map are given approximately equal weights, with the weights set to (4, 3, 3) respectively.

[0123] Subsequently, the system will spatially modulate the preliminary fused weights using the previously generated artifact intensity map to generate the final perceptual saliency map. When the artifact intensity is high, the perceptual saliency of the corresponding region will be appropriately reduced to avoid misjudging noise or artifacts as high perceptual information. Specifically, the system sets an artifact intensity threshold, and when the artifact intensity of a region exceeds this threshold, the perceptual saliency of that region will be gradually attenuated. The attenuation amplitude will become more significant as the artifact intensity increases, but the attenuation process is smooth to avoid abrupt changes. This modulation ensures the accuracy of the final perceptual saliency map and avoids misjudging noise and artifacts as regions of high perceptual saliency. The final generated perceptual saliency map will highlight the truly important, clear, and high-visual-information parts of the image in high-light regions.

[0124] Example Four

[0125] Referring to Figure 3 , this embodiment takes an autonomous driving system as an example to demonstrate the specific process of generating an enhanced image based on frequency domain purification map, spatial domain purification map, time domain stability map and core structure map, using pyramid collaborative fusion algorithm for multi-resolution information fusion, and regulating the fusion weight through saliency map and artifact intensity map.

[0126] The frequency domain purification map generated by the frequency domain decoupling branch (mainly removes periodic ripples, but may retain some random noise and blur), and the spatial domain purification map generated by the spatial domain decoupling branch (mainly removes random noise and blur, but may not completely suppress periodic ripples or slightly smooth details) are respectively constructed into their corresponding Laplacian pyramids. This embodiment constructs a 5-layer Laplacian pyramid, from the coarsest image overview layer (high layer) to the finest detail layer (low layer), each layer contains image detail information at different spatial frequencies.

[0127] The saliency map, artifact intensity map and time domain stability map generated by the time domain decoupling branch (mainly suppresses local distortion and transient high-frequency interference) are respectively constructed into their corresponding Gaussian pyramids. This embodiment constructs a 5-layer Gaussian pyramid corresponding to the number of layers of the Laplacian pyramid, each layer provides perception importance, artifact degree and time domain stability information at the corresponding scale.

[0128] At each level of the pyramid, the system will calculate an initial fusion weight for the corresponding pixel position based on the artifact intensity map and the time domain stability map in the Gaussian pyramid at that level. This initial fusion weight determines the relative contribution of the frequency domain purification map and the spatial domain purification map in the fusion without the intervention of other higher-level perception information. Specifically, for each pixel point, if its value in the artifact intensity map is high (indicating that the noise in this area is serious), the system will tend to give a higher weight to the stronger denoising source (such as the spatial domain purification map, because it has a good suppression effect on random noise). Conversely, if the artifact intensity is low, the weight will be more evenly distributed.

[0129] At the same time, if the pixel point has a high value in the time domain stability map (indicating that the area is more stable in the time domain and has less artifacts), the system will tend to retain its original information or give a higher weight to the source with higher reliability. If the corresponding value in the time domain stability map is low (indicating that the area has flicker or transient interference), the system will tend to rely more on those sources that can effectively suppress such interference.

[0130] The embodiment adopts a linear weighting strategy based on artifact intensity and temporal stability to generate the initial fusion weight. The greater the artifact intensity, the higher the weight of the spatial purification map and the lower the weight of the frequency purification map; the higher the temporal stability, the higher the weight of the current image source. The specific weight distribution is realized through a simple lookup table or threshold judgment, for example, if the artifact intensity exceeds a certain threshold, the initial weight of the spatial purification map increases by 0.2 and the initial weight of the frequency purification map decreases by 0.2. If the temporal stability is lower than a certain threshold, the weights of the two are fine-tuned, and the weight of the more stable source is biased.

[0131] The initial fusion weight calculated above is spatially modulated using the perceptual saliency map in the corresponding level of the Gaussian pyramid to generate the final fusion weight map. In the perceptual salient region, the weight of the more optimal detail image source is enhanced by spatial modulation.

[0132] Specifically, when the pixel value in the perceptual saliency map is high, the system checks the local detail preservation of the frequency purification map and the spatial purification map in that region. By comparing the local gradient amplitudes of the two purification maps in that region, if the edge detail preservation of the frequency purification map is better than that of the spatial purification map, the fusion weight of the frequency purification map in that region will be enhanced. This enhancement is adaptive and will be adjusted according to the level of perceptual saliency. The higher the perceptual saliency, the greater the enhancement. The embodiment adopts a nonlinear modulation function based on the Sigmoid function to map the value of the perceptual saliency map to a modulation factor. When the perceptual saliency is high, the modulation factor will amplify the initial weight of the corresponding image source to ensure that the key details are preserved to the maximum extent.

[0133] The final fusion weight map is used to weight and fuse the frequency purification map and the spatial purification map in the corresponding level of the Laplacian pyramid. For each pixel at each level, the fused Laplacian pyramid value is the weighted sum of the corresponding Laplacian pyramid values of the frequency purification map and the spatial purification map, and the weight is provided by the final fusion weight map. This step organically combines the detail information of the two different denoising results to generate the fused Laplacian pyramid level.

[0134] The preliminary enhanced image has fused the de-noising advantages of each branch and is optimized in terms of perception and artifact level. Finally, the edge of the preliminary enhanced image is enhanced using the core structure map generated in step S2 to generate the final enhanced image. In this embodiment, a non-sharpening mask technique is used, and the core structure map is used as a mask: for the edge region in the core structure map, the system will perform local contrast enhancement on the preliminary enhanced image to further improve the visual clarity and recognition of the image. For example, by subtracting a preliminary enhanced image from a Gaussian blurred version to obtain a detail map, and then using the core structure map as a weight to moderately superimpose the detail map back to the preliminary enhanced image, thereby only sharpening the edge region.

[0135] While embodiments of the present application have been shown and described with reference to the figures, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the spirit and scope of the application, which should be limited only by the scope of the appended claims and their equivalents.

Claims

1. A real-time image quality enhancement method for vehicle camera EMC test, characterized in that, The method comprises the following steps: acquiring real-time image data and historical image data of a vehicle-mounted camera; performing multi-branch collaborative decoupling on the real-time image data; a frequency domain decoupling branch, filtering out EMC noise in the real-time image data in the frequency domain to generate a frequency domain purification image; a structural feature branch, extracting core structures from the frequency domain purification image to generate a core structure image; a spatial domain decoupling branch, under the guidance of the core structure image, performing smoothing processing on random noise of the real-time image data to generate a spatial domain purification image; and a time domain decoupling branch, combining the historical image data and the real-time image data to form time series data, analyzing high-energy pixels, and generating a time domain stability image; performing local image statistical analysis and multi-feature extraction on the real-time image data to generate an artifact intensity image and a perceptual saliency image; based on the frequency domain purification image, the spatial domain purification image, the time domain stability image and the core structure image, performing multi-resolution information fusion by using a pyramid collaborative fusion algorithm, and adjusting and controlling fusion weights by using the artifact intensity image and the perceptual saliency image to generate an enhanced image; wherein, the frequency domain purification image and the spatial domain purification image are respectively constructed into corresponding Laplacian pyramids; the perceptual saliency image, the artifact intensity image and the time domain stability image are respectively constructed into corresponding Gaussian pyramids; based on the artifact intensity image and the time domain stability image in each level of the Gaussian pyramid, initial fusion weights for pixel positions in the corresponding level are calculated; the initial fusion weights are spatially modulated by using the perceptual saliency image in the corresponding level of the Gaussian pyramid to enhance the weight of the image source that is better in retaining details in the perceptual salient region, and a final fusion weight image is generated; the frequency domain purification image and the spatial domain purification image in the corresponding level of the Laplacian pyramid are weighted and fused by using the final fusion weight image to generate a fused Laplacian pyramid level; all the fused Laplacian pyramid levels are combined to generate a preliminary enhanced image; the preliminary enhanced image is edge-enhanced by using the core structure image to generate a final enhanced image; when the pixel value in the perceptual saliency image is high, the local detail retention of the frequency domain purification image and the spatial domain purification image in the region is checked; by comparing the local gradient amplitudes of the two purification images in the region, if the edge detail retention of the frequency domain purification image is better than that of the spatial domain purification image, the fusion weight of the frequency domain purification image in the region is enhanced; a nonlinear modulation function based on a Sigmoid function is used to map the value of the perceptual saliency image to a modulation factor; when the perceptual saliency is high, the modulation factor amplifies the initial weight of the corresponding image source to ensure that the key details are retained to the maximum extent.

2. The real-time image quality enhancement method for EMC test of vehicle-mounted camera according to claim 1, characterized in that, The frequency domain decoupling branch specifically comprises: performing two-dimensional Fourier transform on the real-time image data to obtain a frequency spectrum image; performing peak value detection and conjugate symmetry analysis on the frequency spectrum image, setting a dynamic spectrum threshold according to the energy mean value and standard deviation of the frequency spectrum image, identifying candidate noise frequency points with energy values exceeding the dynamic spectrum threshold and having conjugate symmetry distribution characteristics; and calculating the energy concentration degree of each candidate noise frequency point to dynamically adjust the center frequency and filter bandwidth of the band-stop filter. The band-stop filter is used for filtering candidate noise frequency points, and two-dimensional inverse Fourier transform is performed on all filtered candidate noise frequency points to generate a frequency domain purification map.

3. The real-time image quality enhancement method for EMC testing of vehicle-mounted cameras according to claim 1, characterized in that, The structural feature branch specifically includes: Multi-scale gradient calculation is performed on the frequency domain purification map to generate gradient amplitude maps and gradient direction maps at different scales; non-maximum suppression is performed on the gradient amplitude maps to preliminarily locate edge candidate pixels; the edge candidate pixels with similar gradient directions are connected to generate continuous edge segments; the continuous edge segments with a length less than a preset length threshold and a curvature greater than a preset curvature threshold are removed, and the remaining continuous edge segments after screening are combined to generate a core structure map.

4. The real-time image quality enhancement method for EMC testing of vehicle-mounted cameras according to claim 1, characterized in that, The spatial domain decoupling branch specifically includes: Based on the core structure map, a structure region mask is generated through a morphological dilation operation; the structure region mask is inverted to generate a smooth region mask; based on the structure region mask, bilateral filtering is performed on the real-time image data to generate a first filtering result; Based on the smooth region mask, Gaussian filtering is performed on the real-time image data to generate a second filtering result; The first filtering result and the second filtering result are merged to generate a spatial domain purification map.

5. The real-time image quality enhancement method for EMC testing of vehicle-mounted cameras according to claim 1, characterized in that, The time domain decoupling branch specifically includes: The real-time image data and the historical image data are combined to form time series data; motion estimation based on block matching is performed on the time series data to determine a corresponding matching block for each pixel block in the real-time image data in the historical image data, and a motion vector field is generated; Based on the motion vector field, motion compensation is performed on the historical image data to generate compensated historical image data aligned with the real-time image data; Based on the pixel brightness difference between the real-time image data and the compensated historical image data, a residual sequence is generated; One-dimensional Fourier transform is performed on the residual sequence to obtain a residual frequency spectrum; high-energy pixel coordinates are extracted from the residual frequency spectrum, and a time domain stability map is generated by combining the high-energy pixel coordinates and the real-time image data.

6. The real-time image quality enhancement method for EMC testing of vehicle-mounted cameras according to claim 1, characterized in that, The generation process of the artifact intensity map and the perceptual saliency map specifically includes: Local neighborhood scanning is performed on the real-time image data to calculate pixel value variance and gradient direction histogram entropy of the local neighborhood of a pixel block; the pixel value variance and the gradient direction histogram entropy are weighted and summed to generate the artifact intensity map; parallel feature analysis is performed on the real-time image data, including linear feature analysis using Hough transform, color feature analysis using color range threshold segmentation, and shape feature analysis using contour detection, to generate linear feature maps, color feature maps, and shape feature maps; the linear feature maps, the color feature maps, and the shape feature maps are weighted and fused, and the fusion weights are adjusted using the artifact intensity map to generate a perceptual saliency map.

Images