A target image enhancement method and device with light interference robustness

Abstract

The present application relates to a kind of target image enhancement method and equipment with light interference robustness;The method comprises: obtaining image to be processed;The image is respectively subjected to contrast equalization processing and illumination autonomous compensation processing;And the image data after contrast equalization processing and the image data after illumination autonomous compensation processing are weighted fusion, and preliminary enhancement image is obtained;The preliminary enhancement image is subjected to target region detection, and target region image is obtained;The target region image is input into pre-trained image enhancement model, and the target region image of uniform illumination is generated.The present application solves the problem that the image enhancement method in the prior art is difficult to realize the illumination equalization effect of the whole picture, so that the processed image is locally clear, but the whole image is severely distorted, which cannot be used for subsequent target recognition and other tasks.

Description

Technical Field

[0001] This invention belongs to the field of steel ladle inspection technology, specifically relating to a target image enhancement method and device with robustness against optical interference. Background Technology

[0002] In recent years, image recognition and target detection have been increasingly widely used. However, under complex lighting conditions, the clarity and contrast of images can drop sharply, severely affecting object identification and target detection. For example, the identification of ladle markings and billet numbers in steel mill workshops, and license plate numbers in traffic scenarios, are easily affected by strong and weak light. In the identification of ladle markings and billet numbers or defect detection, performing anti-light interference image enhancement on images of ladle markings, billet numbers, or defect surfaces that are affected by strong or weak light interference before marking or defect detection will greatly improve the accuracy of identification, thereby improving the accuracy of steel coil quality assessment. Similarly, in license plate recognition in traffic scenarios, removing strong light interference from license plates before recognition will have the same effect.

[0003] Traditional image processing methods, such as direct histogram equalization, rely primarily on the image's grayscale statistics, resulting in poor performance under complex lighting conditions. While deep learning-based image enhancement methods can improve the model's robustness to lighting through training, they require substantial amounts of labeled data and computational resources. Furthermore, many traditional image enhancement models often neglect local image characteristics, such as contrast and saturation, which can lead to over-enhancement or distortion. Therefore, existing methods often struggle to achieve adaptive lighting equalization across the entire image; for example, some parts of the processed image may be clearer, but the overall image may be severely distorted, rendering it unusable for subsequent tasks. Thus, a more robust image enhancement method is needed to effectively handle image detection and recognition problems in scenes with significant lighting variations (such as steel stamps and license plate numbers). Summary of the Invention

[0004] Based on the above analysis, the present invention aims to disclose a target image enhancement method and device with robustness against light interference, which solves the problem that existing image enhancement methods are difficult to achieve illumination equalization of the entire image, resulting in locally clear images but severely distorted overall images that cannot be used for subsequent target recognition and other tasks.

[0005] The objective of this invention is mainly achieved through the following technical solutions:

[0006] On one hand, this invention discloses a target image enhancement method with optical interference robustness, comprising:

[0007] Obtain the image to be processed;

[0008] The images are subjected to contrast equalization and autonomous illumination compensation respectively; and the image data after contrast equalization and the image data after autonomous illumination compensation are weighted and fused to obtain a preliminary enhanced image.

[0009] Target region detection is performed on the preliminary enhanced image to obtain the target region image;

[0010] The target region image is input into a pre-trained image enhancement model to generate a target region image with uniform illumination.

[0011] Furthermore, the autonomous illumination compensation process includes:

[0012] Convert the image from RGB space to YUV space;

[0013] Automatic color equalization and tone stretching normalization are performed on the converted YUV image;

[0014] A global exponential transform is performed on the Y channel of the normalized image to adjust the image brightness;

[0015] Convert the YUV image after global exponential transformation back to an RGB image to complete the autonomous illumination compensation operation.

[0016] Furthermore, the automatic color equalization processing includes color difference correction and non-linear contrast enhancement processing of the YUV image, expressed as:

[0017]

[0018] Among them, M c (p)-M c (j) represents the grayscale difference between two pixels p and j, and d(p,j) represents the Euclidean distance between pixels p and j. This indicates non-linear contrast enhancement processing.

[0019] Furthermore, a global exponential transformation is performed using the following formula:

[0020]

[0021] Among them, M Y,, (p) represents the Y-channel grayscale value of pixel p after a global exponential transformation; M Y, (p) represents the grayscale value of the Y channel of pixel p before the global exponential transformation.

[0022] Furthermore, the contrast equalization process includes:

[0023] The image to be processed is divided into multiple sub-regions, and a gray-level distribution histogram is constructed for each sub-region.

[0024] Traverse the grayscale distribution histogram, and for grayscale levels with a pixel count greater than a preset contrast threshold, adjust the corresponding pixel count to the number of pixels corresponding to the contrast threshold.

[0025] The amount of pixels overflowing from each gray level is then evenly distributed across the various gray levels of the histogram.

[0026] After adjusting the grayscale distribution, pixel grayscale values ​​are interpolated at the edge positions of each sub-region to obtain the image after contrast equalization.

[0027] Furthermore, the pixel grayscale value interpolation processing is expressed as follows:

[0028] I R (x,y)=F(I TL (x,y),I TR (x,y),I BL (x,y),I BR (x,y))

[0029] I G (x,y)=F(I TL (x,y),I TR (x,y),I BL (x,y),I BR (x,y))

[0030] I B (x,y)=F(I TL (x,y),I TR (x,y),I BL (x,y),I BR (x,y))

[0031] Where F() is the interpolation function, I TL (x,y), I TR (x,y), I BL (x,y) and I BR (x,y) represents the pixel values ​​of the top-left, top-right, bottom-left, and bottom-right corners of the sub-region where the pixel at position (x,y) is located.

[0032] Furthermore, the weighted fusion of the image data after contrast equalization and the image data after autonomous illumination compensation includes:

[0033] The weighting coefficient for the contrast-equalized image data was set to 0.6, and the weighting coefficient for the image data after autonomous illumination compensation was set to 0.4.

[0034] Based on the weighting coefficients, the image data after contrast equalization and the image data after illumination autonomous compensation are weighted and summed.

[0035] Furthermore, the target region detection in the preliminarily enhanced image includes:

[0036] Based on the pixel value heatmap of the preliminary enhanced image, the confidence level of each pixel as the center point of the image and the offset of each pixel are calculated.

[0037] The center point of the preliminary enhanced image is obtained based on the confidence level;

[0038] Based on the center point and the offset of each pixel, the bounding box of the target region is predicted.

[0039] Furthermore, the training of the image enhancement model includes:

[0040] Multiple uniformly illuminated images and non-uniformly illuminated images are acquired; a first training dataset and a second training dataset are constructed based on the uniformly illuminated images and non-uniformly illuminated images, respectively.

[0041] Images from two training datasets are randomly cropped to obtain images of a preset size;

[0042] A generator and a discriminator are constructed based on the CycleGAN network; the generator is used to learn to convert unevenly lit images in the second training dataset into uniformly lit images in the first training dataset; the discriminator is used to learn to distinguish between real images and images generated by the generator.

[0043] The generator and discriminator are trained using images from the first and second training datasets. After iterative updates of the loss function, the loss is minimized to obtain a converged image enhancement model.

[0044] On the other hand, a computer device is also disclosed, including at least one processor and at least one memory communicatively connected to said processor;

[0045] The memory stores instructions that can be executed by the processor to implement the aforementioned target image enhancement method with optical interference robustness.

[0046] This invention can achieve at least one of the following beneficial effects:

[0047] 1. This method combines autonomous illumination compensation and contrast equalization to coarsely process the image. Through automatic color equalization and global exponential transformation, it effectively enhances the image's color based on global pixels and adjusts the overall brightness and darkness of the image through global exponential transformation, resulting in a brighter image with higher target recognizability. Contrast equalization is then used to smooth local information, adjusting the histogram information of local areas to ensure good local contrast while interpolating between different local areas to maintain continuity and reduce image distortion. The clearer image obtained after coarse processing using these two methods is then used for target region recognition. Finally, a style conversion from uneven to uniform illumination is applied to the target region to enhance target recognition, improving the efficiency and accuracy of target identification.

[0048] 2. This invention employs an innovative method to perform region adaptive filtering on all pixels in each channel of the input image to correct color differences, and through non-linear contrast enhancement, produces a softer contrast enhancement effect, thereby improving the recognizability of target areas in images affected by strong light. Attached Figure Description

[0049] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0050] Figure 1 This is a flowchart of a target image enhancement method with optical interference robustness in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the image processing process of the target image enhancement method with optical interference robustness in an embodiment of the present invention;

[0052] Figure 3 This is a graph showing the loss reduction of the image enhancement model in an embodiment of the present invention; wherein... Figure 3 A represents the generator loss decrease curve; Figure 3 B represents the discriminator loss reduction curve. Detailed Implementation

[0053] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0054] Embodiments of the present invention disclose a target image enhancement method with optical interference robustness, such as... Figure 1 As shown, it includes the following steps:

[0055] Step S1: Obtain the image to be processed;

[0056] Specifically, the images to be processed can be images captured by a camera or images obtained by segmenting video files; for example, images of ladle marking data with strong or weak light interference obtained by segmenting video files from steel mills, and license plate images with strong and weak light interference.

[0057] Step S2: Perform contrast equalization and illumination autonomous compensation processing on the acquired images to be processed; and perform weighted fusion on the image data after contrast equalization processing and the image data after illumination autonomous compensation processing to obtain a preliminary enhanced image.

[0058] Specifically, such as Figure 2 As shown, autonomous illumination compensation processing includes:

[0059] Convert the image from RGB space to YUV space;

[0060] Automatic color equalization and tone stretching normalization are performed on the converted YUV image. The tone values ​​in the YUV image are linearly transformed and mapped to the range [0,1] to achieve tone stretching normalization. Max-min normalization is used here.

[0061] A global exponential transform is performed on the Y channel of the normalized image to adjust the image brightness;

[0062] Convert the YUV image after global exponential transformation back to an RGB image to complete the autonomous illumination compensation operation.

[0063] More specifically, color image color spaces include RGB, YUV, and LAB spaces. RGB space is a color model that represents color as three components: red (R), green (G), and blue (B). In RGB space, the color of each pixel is represented by the values ​​of these three components, each ranging from 0 to 255. Different combinations of these component values ​​result in different colors. YUV space is a color model that represents color as two components: lightness (Y) and chromaticity (UV). LAB space is a color model that represents color as lightness (L) and two chromaticity components (A and B). When adjusting color and lightness in the traditional RGB space, the lightness and color information are intertwined, causing adjustments to each other. Comparative experiments show that processing in RGB space leads to color distortion, requiring separate adjustments for either lightness or color information. Therefore, RGB images can be converted to YUV or LAB space for autonomous illumination compensation. Comparative experiments show that processing in YUV space yields better results.

[0064] The purpose of this embodiment's autonomous illumination compensation is to adjust image brightness, making it uniform while maintaining color balance and preventing distortion. Therefore, the image is first converted from RGB space to YUV space to obtain a YUV space image M. yuv (x,y,c), that is:

[0065]

[0066] Among them, M yuv (x,y,c) represents the image converted to YUV space; I R (x,y), I G (x,y), I B (x,y) represent the pixel values ​​at (x,y) positions in the R, G, and B channels of the RGB image, respectively; c = 0, 1, 2, representing the Y, U, and V channels, respectively.

[0067] After converting the image to YUV space, automatic color equalization is first performed to maintain color balance. This includes color difference correction and non-linear contrast enhancement of the YUV image, as shown below:

[0068]

[0069] Among them, M c (p)-M c (j) represents the grayscale difference between two pixels p and j, and d(p,j) represents the Euclidean distance between pixels p and j, used to express the influence weight of different pixels on pixel p. This indicates nonlinear contrast enhancement processing to provide a nonlinear contrast enhancement effect, resulting in a smoother contrast enhancement effect; c = 0, 1, 2, representing the Y, U, and V channels respectively; this embodiment uses a continuous and differentiable nonlinear function for contrast enhancement processing, which has smooth characteristics, which helps to reduce abrupt gradient changes and make the filtering results smoother; and through experiments, it can be seen that the peak signal-to-noise ratio (PSNR) after image processing can be improved by 1.4 compared with conventional methods, and the structural similarity (SSIM) can be improved by 0.03, that is, the loss before and after image processing is small, the structural similarity is high, and the overall effect is significantly improved.

[0070] The automatic color equalization algorithm considers the spatial relationship between color and brightness in an image, performing adaptive filtering of local characteristics to achieve brightness, color, and contrast adjustments for images with local and non-linear features, while satisfying the gray world theory assumption and the white spot assumption. The value and correction of each pixel in the image depend on its surrounding pixels; pixels closer to the pixel have a greater impact, while pixels farther away have a smaller impact, resulting in a smoother and softer image effect.

[0071] Furthermore, the image after automatic color equalization is subjected to tone stretching and normalization to map the image pixel values ​​to the range [0,1], thus obtaining the final automatic color equalization result image.

[0072] To prevent certain local areas of the image from being too bright or too dark, this embodiment also employs a global exponential transformation to perform secondary processing on the image. This transforms the pixel values ​​based on the brightness values ​​of different regions in the image, preventing local overexposure or underexposure. This transformation can perform a non-linear transformation on the pixel values ​​in the image to adjust the image's exposure, improve the image's visual effect, and correct images with excessively high or low gray levels, thereby enhancing contrast.

[0073] The global exponential transformation formula is expressed as follows:

[0074]

[0075] Among them, M Y,, (p) represents the Y-channel grayscale value of pixel p after global exponential transformation; M Y, (p) represents the Y-channel grayscale value of pixel p before global exponential transformation.

[0076] That is, if the gray value of the Y channel at the current position is greater than 150, an exponential transformation with a value of 1.1 is performed to adjust the brightness and prevent the image from being too bright; if the Y channel value at the current position is not greater than 53, an exponential transformation with a value of 0.8 is performed to increase the brightness to a certain extent; finally, an exponential transformation with a value of 0.8 is performed on the entire image to increase the overall image brightness.

[0077] In the ladle workshop applied in this embodiment, when the gray value of the Y channel is greater than 150 or less than or equal to 53, it indicates that the current pixel is either too bright or too dark, requiring brightness adjustment. This is achieved through exponential transformation with specific values. Extensive experimental verification shows that this embodiment first applies an exponential transformation of 1.1 to pixels with a Y channel gray value greater than 150. This causes pixels with higher brightness values ​​to increase faster relative to pixels with lower brightness values, thereby reducing detail loss in high-brightness areas, making the image appear more balanced, with higher SSIM and PSNR, less image loss, and preventing the ladle area image from becoming too dark. For pixels with a Y channel gray value no greater than 53, an exponential transformation of 0.8 is applied, causing pixels with lower brightness values ​​to increase faster relative to pixels with higher brightness values. This enhances detail in low-brightness areas, making the image appear brighter, with higher SSIM and PSNR, less image loss, and preventing the ladle area image from becoming too bright. Finally, an exponential transformation of 0.8 is applied to adjust the overall image, increasing its overall brightness. This global exponential transformation, as described in this embodiment, allows for the adjustment of brightness distribution while preserving image details, resulting in a more balanced and brighter image with reduced overly bright areas. This processing improves the visual effect of the image, making it more consistent with human visual perception.

[0078] Finally, the YUV space image after global exponential transformation is converted back to an RGB space image. The resulting RGB image is clearly visible in certain areas, and the visual effect is good.

[0079] Furthermore, such as Figure 2 As shown, the contrast equalization process includes:

[0080] The image to be processed is divided into multiple sub-regions, and a gray-level distribution histogram is constructed for each sub-region.

[0081] Traverse the grayscale distribution histogram, and for grayscale levels with a pixel count greater than a preset contrast threshold, adjust the corresponding pixel count to the number of pixels corresponding to the contrast threshold.

[0082] The amount of pixels overflowing from each gray level is then evenly distributed across the various gray levels of the histogram.

[0083] After adjusting the grayscale distribution, pixel grayscale values ​​are interpolated at the edge positions of each sub-region to obtain the image after contrast equalization.

[0084] Specifically, the degree of contrast enhancement is controlled by cropping and redistributing the histogram, including:

[0085] Perform grayscale analysis on each sub-region, that is, count the grayscale levels of all pixels in each sub-region, and construct the grayscale distribution map of the region based on the count data;

[0086] Histogram limit adjustment sets a maximum allowed pixel count for the grayscale distribution of each sub-region. If a grayscale level exceeds this limit, its pixel count is reduced to the set upper limit.

[0087] Pixel redistribution involves reducing the number of pixels at excessive gray levels and then evenly distributing the reduced number of pixels across the various gray levels of the histogram to achieve smooth contrast adjustment and avoid abrupt transitions.

[0088] Finally, grayscale interpolation is used to integrate the histograms of each redistributed sub-region, thus forming the brightness adjustment map for the entire image. This embodiment compares the use of bilinear interpolation and bicubic interpolation. Bicubic interpolation is more complex and slower, making it unsuitable for industrial environments; therefore, bilinear interpolation is used, expressed as:

[0089]

[0090] Where F() is the interpolation function, I TL (x,y), I TR (x,y), I BL (x,y) and I BR (x,y) represents the pixel values ​​of the top-left, top-right, bottom-left, and bottom-right corners of the sub-region where the pixel at position (x,y) is located.

[0091] Finally, the three channels are merged to obtain the multi-channel graph M. rgb (x,y,c), where c represents the three-channel image of 0, 1, and 2 (corresponding to Y, U, and V), i.e.:

[0092]

[0093] In this embodiment, contrast equalization ensures that the enhancement of contrast within each sub-region is limited, and interpolation techniques are used to ensure that these limited enhancements produce a smooth transition across the entire image. This achieves contrast enhancement without causing over-enhancement or distortion.

[0094] Furthermore, the weighting coefficient of the contrast equalization processed image data is set to 0.6, and the weighting coefficient of the image data after autonomous illumination compensation is set to 0.4. Based on the weighting coefficients, the image data after contrast equalization and the image data after autonomous illumination compensation are weighted and summed to obtain a preliminary enhanced image combining contrast equalization and autonomous illumination compensation.

[0095] This embodiment employs an autonomous illumination compensation method, based on automatic color equalization and global exponential transformation. Automatic color equalization effectively enhances image color by adjusting the current pixel based on global pixels, while global exponential transformation effectively adjusts the overall brightness of the image, resulting in a brighter overall image. Therefore, this method offers good visual effects and high object recognition, but may distort local information. Intelligent contrast equalization effectively smooths local information, adjusting the histogram information of local areas to ensure good local contrast while interpolating between different local areas to ensure good continuity and low image distortion. However, its clarity is not as good as the autonomous illumination compensation method. Therefore, combining these two methods ensures low image distortion and a clearer image.

[0096] Step S3: Detect the target region in the preliminary enhanced image to obtain the target region image;

[0097] Specifically, the CenterNet method can be used for target region detection. First, based on the pixel value heatmap of the preliminary enhanced image, the confidence level of each pixel as the image center point and the offset of each pixel are calculated. Then, the center point of the preliminary enhanced image is obtained based on the confidence level. Based on the center point and the offset of each pixel, the bounding box of the target region is predicted.

[0098] It should be noted that the preliminary enhanced image after combining contrast equalization and autonomous illumination compensation is a coarse image processing step. The image is relatively clear and can be used to accurately identify the target detection region of interest. In the steel number or license plate recognition field of this embodiment, the regions of interest are only the steel number portion of the ladle and the license plate number portion of the license plate. Therefore, by performing target region detection such as steel number or license plate number on the preliminary enhanced image obtained after coarse image processing, the target region image to be processed is segmented and used for subsequent fine image processing for image enhancement. This avoids performing image enhancement processing on the entire image, thus improving the efficiency of image enhancement and the accuracy of subsequent steel number or license plate number recognition.

[0099] Step S4: Input the target region image into a pre-trained image enhancement model to generate a target region image with uniform illumination.

[0100] Specifically, the image enhancement model is trained using the following method:

[0101] Multiple uniformly illuminated images and non-uniformly illuminated images are acquired; a first training dataset and a second training dataset are constructed based on the uniformly illuminated images and non-uniformly illuminated images, respectively.

[0102] Images from two training datasets are randomly cropped to obtain images of a preset size;

[0103] A generator and a discriminator are constructed based on the CycleGAN network; the generator is used to learn to convert unevenly lit images in the second training dataset into uniformly lit images in the first training dataset; the discriminator is used to learn to distinguish between real images and images generated by the generator.

[0104] The generator and discriminator are trained using images from the first and second training datasets. After iterative updates of the loss function, the loss is minimized to obtain a converged image enhancement model.

[0105] More specifically, CycleGAN generative adversarial network can transform the style of an image from unevenly lit to uniformly lit without a large amount of labeled information. It is an unsupervised learning method that can perform image transformation even when there is a lack of paired training data and can be trained on two unrelated sets of images without the need for one-to-one image pairs.

[0106] First, using the aforementioned images of the iron bag label and license plate, a training dataset is created. The training dataset contains two datasets, A and B, which respectively store images with uneven lighting and images with uniform lighting. However, the datasets in A and B do not need to correspond one-to-one.

[0107] Images in the training dataset are randomly cropped to obtain images of a specified size (e.g., 256*256) for training, which increases the network's generalization ability and prevents overfitting.

[0108] Using the Adam optimizer, the CycleGAN network was trained for 200 epochs. Visdom was used to record the generator loss decrease curve, discriminator loss decrease curve, and cycle consistency loss curve, respectively. Figure 3 A and Figure 3 As shown in Figure B, the generator loss curve and the discriminator loss curve are shown respectively; after iterative training, a converged image enhancement model is obtained.

[0109] Indicators for resisting illumination interference include average gradient, information entropy, PSNR (peak signal-to-noise ratio), and SSIM (structural similarity).

[0110] The two metrics, average gradient and information entropy, aim to ensure that the enhanced image contains sufficient image information to facilitate subsequent detection and recognition tasks. Higher average gradient and information entropy are generally better, indicating a greater amount of image information. The average gradient is calculated by first performing gradient calculations on the image (using the Sobel operator), then calculating the gradient magnitude of each pixel, and finally averaging the gradient magnitudes of all pixels to obtain the average gradient.

[0111] PSNR stands for Peak Signal-to-Noise Ratio, which measures the degree of quality loss before and after image processing. PSNR is usually expressed in decibels (dB). SSIM stands for Structural Similarity Index, which is a metric for measuring the similarity between two images. A higher PSNR indicates less loss after image processing; a higher SSIM indicates better structural similarity before and after image processing.

[0112] To verify the effectiveness of the image enhancement method of this invention, this embodiment uses ladle label data. A video file from a steel plant is segmented into frames to obtain images. Images of ladle label data with strong or weak light interference are extracted from these images. Simultaneously, license plate images with strong and weak light interference from the CCPD_Green training set in the open-source dataset CCPD 2019 are used for image enhancement method verification. The ladle images with strong or weak light interference and the license plate photos with strong light interference are categorized into license plate with strong light interference, ladle 1 with strong light interference, ladle 2 with strong light interference, and ladle 3 with weak light interference for experiments. The method of this embodiment is compared with automatic color equalization, intelligent contrast equalization, and MSRCR (multi-scale residual enhancement) methods on objective evaluation metrics. The results are shown in Tables 1, 2, 3, and 4 below.

[0113] Table 1. Results of different models for handling license plates affected by strong light interference.

[0114]

[0115] Table 2. Results of different models on the processing of iron bag 1 under strong light interference.

[0116]

[0117] Table 3. Results of different models on the treatment of strong light interference in iron bag 2

[0118]

[0119] Table 4 shows the processing results of different models for iron bag 3 under low light interference.

[0120]

[0121] The experimental results show that the image enhancement method of this invention outperforms other existing image enhancement methods in terms of average gradient, information entropy, PSNR (peak signal-to-noise ratio), and SSIM (structural similarity). The processed target region has high clarity and can be used for target recognition such as steel number and license plate number. The image processed by the method of this invention achieves a balance between subjective and objective indicators, with good intuitive effect and objective indicators and low distortion.

[0122] Another embodiment of the present invention also discloses a computer device, including at least one processor and at least one memory communicatively connected to said processor;

[0123] The memory stores instructions that can be executed by the processor to implement the aforementioned target image enhancement method with optical interference robustness.

[0124] In summary, such as Figure 2 As shown, the target image enhancement method of this invention, which is robust to light interference, combines autonomous illumination compensation and contrast equalization methods for coarse processing of the image to be processed. Through automatic color equalization and global exponential transformation, the image color is effectively enhanced based on global pixels, and the overall brightness and darkness relationship of the image is effectively adjusted through global exponential transformation, making the entire image brighter and resulting in an image with high target recognizability. The contrast equalization method is used to smooth local information, adjusting the histogram information of local areas to ensure good local contrast while interpolating between different local areas to ensure continuity between regions and reduce image distortion. The clearer image obtained after coarse processing using the two methods is then used for target region recognition. Furthermore, a style conversion from uneven illumination to uniform illumination is performed on the target area to enhance target recognition, improving the efficiency and accuracy of target recognition. This invention employs an innovative method to perform region-adaptive filtering on all pixels in each channel of the input image for color difference correction, and through non-linear contrast enhancement, produces a softer contrast enhancement effect, improving the target region recognizability of images affected by strong light.

[0125] Those skilled in the art will understand that all or part of the methods in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0126] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A target image enhancement method with robustness against optical interference, characterized in that, include: Obtain the image to be processed; The images are subjected to contrast equalization and autonomous illumination compensation respectively. The image data after contrast equalization and the image data after autonomous illumination compensation are weighted and fused to obtain a preliminary enhanced image. The autonomous illumination compensation process includes: converting the image from RGB space to YUV space; performing automatic color equalization and tone stretching normalization on the converted YUV image; performing a global exponential transformation on the Y channel of the normalized image to adjust the image brightness; and converting the YUV image after the global exponential transformation back to an RGB image to complete the autonomous illumination compensation operation. The global exponent transformation is performed using the following formula: in, For pixels p The grayscale value of the Y channel after global exponential transformation; For pixels p The grayscale value of the Y channel before global exponential transformation; The target region image is obtained by performing target region detection on the preliminary enhanced image; The target region image is input into a pre-trained image enhancement model to generate a target region image with uniform illumination.

2. The target image enhancement method with optical interference robustness according to claim 1, characterized in that, The automatic color equalization processing includes color difference correction and nonlinear contrast enhancement processing of the YUV image, expressed as: in, For two pixels p and j The difference in grayscale values, Represents pixels p and j European distance, This indicates non-linear contrast enhancement processing; , representing the Y, U and V channels respectively.

3. The target image enhancement method with optical interference robustness according to claim 1, characterized in that, The contrast equalization process includes: The image to be processed is divided into multiple sub-regions, and a gray-level distribution histogram is constructed for each sub-region. Traverse the grayscale distribution histogram, and for grayscale levels with a pixel count greater than a preset contrast threshold, adjust the corresponding pixel count to the number of pixels corresponding to the contrast threshold. The amount of pixels overflowing from each gray level is then evenly distributed across the various gray levels of the histogram. After adjusting the grayscale distribution, pixel grayscale values ​​are interpolated at the edge positions of each sub-region to obtain the image after contrast equalization.

4. The target image enhancement method with optical interference robustness according to claim 3, characterized in that, The pixel grayscale value interpolation process is represented as follows: ； in, For interpolation functions, , , and for The pixel values ​​of the top-left, top-right, bottom-left, and bottom-right corners of the sub-region where the location pixel is located.

5. The target image enhancement method with optical interference robustness according to claim 1, characterized in that, The weighted fusion of image data after contrast equalization and image data after autonomous illumination compensation includes: The weighting coefficient for the contrast-equalized image data was set to 0.6, and the weighting coefficient for the image data after autonomous illumination compensation was set to 0.

4. Based on the weighting coefficients, the image data after contrast equalization and the image data after illumination autonomous compensation are weighted and summed.

6. The target image enhancement method with optical interference robustness according to claim 1, characterized in that, The target region detection in the preliminarily enhanced image includes: Based on the pixel value heatmap of the preliminary enhanced image, the confidence level of each pixel as the center point of the image and the offset of each pixel are calculated. The center point of the preliminary enhanced image is obtained based on the confidence level; Based on the center point and the offset of each pixel, the bounding box of the target region is predicted.

7. The target image enhancement method with optical interference robustness according to claim 1, characterized in that, The training of the image enhancement model includes: Multiple uniformly illuminated images and non-uniformly illuminated images are acquired; a first training dataset and a second training dataset are constructed based on the uniformly illuminated images and non-uniformly illuminated images, respectively. Images from two training datasets are randomly cropped to obtain images of a preset size; A generator and a discriminator are constructed based on the CycleGAN network; the generator is used to learn to convert unevenly lit images in the second training dataset into uniformly lit images in the first training dataset; the discriminator is used to learn to distinguish between real images and images generated by the generator. The generator and discriminator are trained using images from the first and second training datasets. After iterative updates of the loss function, the loss is minimized to obtain a converged image enhancement model.

8. A computer device, characterized in that, It includes at least one processor and at least one memory communicatively connected to the processor; The memory stores instructions executable by the processor to implement the target image enhancement method with optical interference robustness as described in any one of claims 1-7.

Images