Night image enhancement and noise suppression integrated system based on deep learning
Through the night image enhancement and noise suppression system based on deep learning, the convolutional neural network and noise suppression algorithm are used to solve the problem of low image quality in low light environments, and the image brightness and contrast improvement and noise removal are achieved, suitable for real-time Application scenario.
Patent Information
- Application Number
- CN202411010405.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-12
AI Technical Summary
Images acquired in low-light environments often have insufficient brightness, low contrast and noise interference, resulting in significant decline in image quality. The prior art image enhancement and noise suppression effects are limited and the processing speed is slow.
Adopt a night image enhancement and noise suppression integrated system based on deep learning, including image preprocessing module, deep learning model module and noise suppression module, and use convolutional neural networks, generative adversarial networks or variational autoencoders to automatically learn image features and adaptive enhancement, combined with a noise suppression algorithm to remove image noise, and finally adjust brightness, contrast and color through the post-processing module to improve image quality.
It significantly improves the brightness, contrast and detail clarity of night images, effectively removes noise, improves processing efficiency, is suitable for real-time application scenarios, and ensures that the visual effect of the output image is natural and beautiful.
Smart Images

Figure BDA0004964170510000021 
Figure BDA0004964170510000091 
Figure FDA0004964170500000011
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field related to image processing, and specifically relates to an integrated system for nighttime image enhancement and noise suppression based on deep learning. Background Art
[0002] Images acquired in low-light environments often have problems such as insufficient brightness, low contrast, and noise interference, resulting in a significant decrease in image quality. These problems are particularly prominent in the fields of surveillance, transportation, medical imaging, and night photography. Traditional image enhancement technology and noise suppression algorithms have limited effects and slow processing speeds in enhancing image details and reducing noise. Therefore, a new image enhancement and noise suppression system is needed to improve the quality of nighttime images. Summary of the invention
[0003] The purpose of the present invention is to provide an integrated system for night image enhancement and noise suppression based on deep learning, so as to solve the problem raised in the above background technology that images acquired in low-light environments often have problems such as insufficient brightness, low contrast and noise interference, resulting in a significant decrease in image quality.
[0004] To achieve the above object, the present invention provides the following technical solutions:
[0005] An integrated system for nighttime image enhancement and noise suppression based on deep learning. The system includes
[0006] Image preprocessing module: normalizes the input nighttime images to improve the accuracy of subsequent processing;
[0007] Deep learning model module:
[0008] Using convolutional neural networks, we designed a deep learning model specific to nighttime image processing. The model can automatically learn image features and perform adaptive enhancement.
[0009] Models that use generative adversarial networks or variational autoencoders to enhance image details and brightness;
[0010] Noise Suppression Module:
[0011] Combined with the output of the image enhancement process, the corresponding noise suppression algorithm is designed to effectively remove random noise in the image;
[0012] Achieve adaptive processing of different noise types and improve the generalization ability of the system.
[0013] Post-processing module: adjusts the brightness, contrast and color of the enhanced image to ensure the visual naturalness and beauty of the final output image.
[0014] The image preprocessing module includes the following:
[0015] S1. Normalization processing:
[0016] Normalization is an important step in image preprocessing. Its main purpose is to convert the input image data into a uniform range, [0, 1] or [-1, 1]. Through normalization, the difference between images under different lighting conditions can be reduced, thereby improving the accuracy of subsequent processing. The following methods can be used for specific implementation:
[0017] Linear Normalization:
[0018] Linearly map the pixel values of the input image to a certain range. The formula is:
[0019]
[0020] Where I is the original pixel value, I_{\text{min}} and I_{\text{max}} are the minimum and maximum pixel values of the image, respectively, I' is the normalized pixel value, and Z-score standardization: the mean and standard deviation of each channel of the image are standardized to eliminate the differences in illumination and contrast between different images:
[0021] I'=\frac{I-\mu}{\sigma}
[0022] Among them, \mu is the mean of the image pixels, \sigma is the standard deviation;
[0023] S2. Noise removal: Before normalizing the image, some simple filters can be used to reduce the noise in the image, which helps improve the effect of subsequent steps;
[0024] Noise removal methods include the following: mean filtering: taking the average of the pixels around each pixel in the image to reduce random noise; median filtering: effectively removing salt and pepper noise by replacing the pixel value with the median of its neighborhood; bilateral filtering: taking into account the spatial distance between pixels and the similarity of pixel values, so as to better retain edge information while removing noise;
[0025] S3, Image enhancement: After normalization, some image enhancement techniques can be applied to improve the visual effect of the image, especially in low-light environments;
[0026] The enhancement methods include: histogram equalization: by adjusting the histogram of the image, the pixel distribution of the image is made more uniform and the contrast is improved; gamma transform: used to adjust the brightness of the image and increase the brightness of the darker parts of the image through nonlinear transformation; Laplace filtering: enhances the edge details of the image and improves the clarity of the image.
[0027] Deep Learning Model Module
[0028] In the deep learning model of night image processing, the combination of convolutional neural network (CNN) and generative adversarial network (GAN) or variational autoencoder (VAE) is used to automatically learn image features and perform adaptive enhancement, which is effective as follows:
[0029] S1. Convolutional Neural Network (CNN)
[0030] Network architecture design:
[0031] Input layer: receives normalized night images with a size of 224x224 or other standard sizes;
[0032] Convolutional layer: Use multiple convolutional layers to extract image features. The convolution kernel size can be 3x3 or 5x5. By increasing the number of convolution kernels layer by layer (such as 16, 32, 64, etc.), more advanced features can be extracted.
[0033] Activation function: ReLU activation function (Rectified Linear Unit) is used to increase the nonlinear expression ability of the model;
[0034] Pooling layer: Use the Max Pooling layer to downsample the image feature map, reduce the number of parameters, and improve the stability of the features;
[0035] Feature extraction:
[0036] The convolution layer learns features, and each convolution kernel can be automatically optimized during the training process to extract features such as edges, textures, and shapes in the image. In the later stage of the model, a fully connected layer can be used to map the features to the output layer to achieve the final image classification or enhancement task;
[0037] S2. Generative Adversarial Network (GAN)
[0038] GAN Architecture:
[0039] Generator:
[0040] Input: A random noise vector in the latent space (e.g. a 100-dimensional vector).
[0041] Network structure: It consists of multiple fully connected layers and transposed convolutional layers to transform the latent vector into an enhanced night image;
[0042] Output: A network that generates close to real (enhanced) images;
[0043] Discriminator:
[0044] Input: real augmented images and images generated by the generator;
[0045] Network structure: Use convolutional layers to extract features, and finally use the Sigmoid activation function to output the true probability of the image (0 for false, 1 for true).
[0046] Training process:
[0047] GAN training is an adversarial process. The generator competes with the discriminator. The generator strives to improve the authenticity of the image, while the discriminator strives to distinguish between real and fake images. After many iterations, the generator can generate high-quality enhanced images with significantly improved detail restoration capabilities.
[0048] S3. Variational Autoencoder (VAE)
[0049] VAE Architecture:
[0050] Encoder:
[0051] Network structure: Using convolutional layers, the input nighttime image is encoded into distribution parameters (mean and variance) of the latent space;
[0052] Decoder:
[0053] Network structure: Use the transposed convolution layer to reconstruct the points in the latent space back to the space of the original image, with the goal of generating an enhanced image;
[0054] Training objectives:
[0055] VAE optimizes reconstruction loss and Kullback-Leibler divergence (KL divergence) during training, and can learn the potential representation of images, so that the model can not only reconstruct the input image, but also generate new high-quality images through the latent space;
[0056] S4. Comprehensive application
[0057] Combining CNN with GAN or VAE can form an efficient deep learning model for night image enhancement:
[0058] First, CNN is used to extract the features of night images;
[0059] Then, GAN or VAE is used to generate and enhance the image to improve details and brightness.
[0060] In the process of image enhancement, noise suppression is an important step, which can significantly improve the visual quality of the image and the reliability of subsequent processing. Specifically, it is Stirk filtering. The steps are as follows:
[0061] Fourier transform: Perform a Fourier transform on an image to convert the image from the spatial domain to the frequency domain.
[0062] Low-pass filter design: Design a Nyquist low-pass filter, select an appropriate cutoff frequency, and remove high-frequency components above this frequency.
[0063] Frequency domain multiplication: Multiply the low-pass filter by the Fourier transformed image element-wise.
[0064] Inverse transform: The result is converted back to the spatial domain through inverse Fourier transform to obtain the denoised image.
[0065] In the post-processing module of image processing, it is very important to adjust the brightness, contrast and color of the enhanced image to ensure that the final output image has a natural and beautiful visual effect. The specific steps are as follows:
[0066] S1. Adjust brightness
[0067] Here’s how:
[0068] Linear adjustment: Increase or decrease the brightness of all pixels by simply adding a constant value (brightness increment). For each pixel value I(x,y), the adjusted new value is:
[0069] I′(x,y)=I(x,y)+δ
[0070] Among them, \delta is the adjusted brightness value, and it is necessary to ensure that the new value does not exceed the valid range (for example, 0-255). Gamma transformation: Gamma transformation can adjust the brightness nonlinearly:
[0071] I'(x,y)=255\cdot\left(\frac{I(x,y)}{255}\right)^{\gamma}
[0072] Among them, \gamma is the gamma value, \gamma<increases brightness, \gamma> reduces brightness,
[0073] S2. Adjust contrast: Contrast adjustment is used to enhance the difference between light and dark areas in an image and improve the visual hierarchy. The specific methods are as follows:
[0074] Simple linear contrast adjustment:
[0075] I'(x,y)=\alpha\cdot(I(x,y)-128)+128
[0076] Among them, \alpha>1 enhances the contrast, and it is necessary to ensure that the new value is still within the valid range;
[0077] Histogram equalization: It can enhance the contrast by changing the gray value distribution of the image, which is especially suitable for low-contrast images.
[0078] Adjust color: Color adjustment includes adjusting saturation and hue to ensure that the color is natural and attractive. The method is as follows:
[0079] Saturation adjustment: You can increase or decrease the saturation by mixing the color value with the grayscale image. Saturation adjustment can specify a ratio:
[0080] I'(x,y)=I(x,y)+S\cdot(I(x,y)-I_{gray}(x,y))
[0081] Where I_{gray} is the gray value of the same pixel;
[0082] Hue adjustment: Change the overall hue of the image by rotating the hue ring and using the HSL model for processing;
[0083] Overall image adjustment steps:
[0084] Brightness adjustment: Improve the brightness of the image as needed;
[0085] Contrast enhancement: Enhance the contrast by linear transformation or histogram equalization;
[0086] Color correction: increase saturation and adjust hue to ensure natural colors;
[0087] Visual evaluation: Perform visual checks after processing to ensure that the image maintains a natural appearance for the user’s aesthetic experience;
[0088] Final output: After completing the above adjustments, you can use the image output function to save the image, output the final result in a common format, and ensure the quality of the output image. Through these steps, the post-processing module can effectively improve the visual effect of the image, ensuring that the enhanced image is not only clear and visible, but also visually natural and beautiful.
[0089] Compared with the prior art, the present invention provides an integrated system for nighttime image enhancement and noise suppression based on deep learning, which has the following beneficial effects:
[0090] Significantly improved image quality: Through the automatic feature learning and adaptive enhancement of the deep learning model, the visual quality of night images, such as brightness, contrast, and detail clarity, has been significantly improved, and clear images can be obtained even in extremely low light conditions;
[0091] Effective noise suppression: The noise suppression algorithm used by the system can effectively remove random noise in the image, such as Gaussian noise, salt and pepper noise, etc., keep the image details clear, and avoid the impact of noise on image quality;
[0092] High processing efficiency: The high-efficiency processing capability of the deep learning model enables the system to complete image processing in a relatively short time, which is suitable for real-time or near-real-time application scenarios such as nighttime monitoring and autonomous driving.
[0093] Strong adaptability: Whether it is the adaptive image generation of the Generative Adversarial Network (GAN) or the learning potential representation of the Variational Autoencoder (VAE), the system can automatically adjust the processing strategy according to different nighttime environments and image contents, improving the generalization ability of image enhancement;
[0094] Post-processing optimization: By adjusting brightness, contrast and color, the post-processing module can ensure that the enhanced image is visually natural and beautiful, meeting the visual needs of different scenarios;
[0095] Improved user experience: High-quality night images are not only important in professional fields such as security monitoring and night photography, but also can improve the nighttime visual experience of ordinary users, such as nighttime navigation and image capture in nighttime outdoor activities.
[0096] Technological integration and innovation: The system integrates multiple technologies such as deep learning, image processing, and signal processing, reflecting the integration and innovation of interdisciplinary technologies and providing new ideas and directions for the development of the field of night image processing. DETAILED DESCRIPTION
[0097] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0098] The present invention provides a technical solution:
[0099] An integrated system for nighttime image enhancement and noise suppression based on deep learning. The system includes
[0100] Image preprocessing module: normalizes the input nighttime images to improve the accuracy of subsequent processing;
[0101] Deep learning model module:
[0102] Using convolutional neural networks, we designed a deep learning model specific to nighttime image processing. The model can automatically learn image features and perform adaptive enhancement.
[0103] Models that use generative adversarial networks or variational autoencoders to enhance image details and brightness;
[0104] Noise Suppression Module:
[0105] Combined with the output of the image enhancement process, design corresponding noise suppression algorithms (such as Nyquist filtering, bilateral filtering, etc.) to effectively remove random noise in the image;
[0106] Achieve adaptive processing of different noise types and improve the generalization ability of the system.
[0107] Post-processing module: adjusts the brightness, contrast and color of the enhanced image to ensure the visual naturalness and beauty of the final output image.
[0108] The image preprocessing module includes the following:
[0109] S1. Normalization processing:
[0110] Normalization is an important step in image preprocessing. Its main purpose is to convert the input image data into a uniform range, [0, 1] or [-1, 1]. Through normalization, the difference between images under different lighting conditions can be reduced, thereby improving the accuracy of subsequent processing. The following methods can be used for specific implementation:
[0111] Linear Normalization:
[0112] Linearly map the pixel values of the input image to a certain range. The formula is:
[0113]
[0114] Where I is the original pixel value, I_{\text{min}} and I_{\text{max}} are the minimum and maximum pixel values of the image, respectively, I' is the normalized pixel value, and Z-score standardization: the mean and standard deviation of each channel of the image are standardized to eliminate the differences in illumination and contrast between different images:
[0115] I'=\frac{I-\mu}{\sigma}
[0116] Among them, \mu is the mean of the image pixels, \sigma is the standard deviation;
[0117] S2. Noise removal: Before normalizing the image, some simple filters can be used to reduce the noise in the image, which helps improve the effect of subsequent steps;
[0118] Noise removal methods include the following: mean filtering: taking the average of the pixels around each pixel in the image to reduce random noise; median filtering: effectively removing salt and pepper noise by replacing the pixel value with the median of its neighborhood; bilateral filtering: taking into account the spatial distance between pixels and the similarity of pixel values, so as to better retain edge information while removing noise;
[0119] S3, Image enhancement: After normalization, some image enhancement techniques can be applied to improve the visual effect of the image, especially in low-light environments;
[0120] The enhancement methods include: histogram equalization: by adjusting the histogram of the image, the pixel distribution of the image is made more uniform and the contrast is improved; gamma transform: used to adjust the brightness of the image and increase the brightness of the darker parts of the image through nonlinear transformation; Laplace filtering: enhances the edge details of the image and improves the clarity of the image.
[0121] Deep Learning Model Module
[0122] In the deep learning model of night image processing, the combination of convolutional neural network (CNN) and generative adversarial network (GAN) or variational autoencoder (VAE) is used to automatically learn image features and perform adaptive enhancement, which is effective as follows:
[0123] S1. Convolutional Neural Network (CNN)
[0124] Network architecture design:
[0125] Input layer: receives normalized night images with a size of 224x224 or other standard sizes;
[0126] Convolutional layer: Use multiple convolutional layers to extract image features. The convolution kernel size can be 3x3 or 5x5. By increasing the number of convolution kernels layer by layer (such as 16, 32, 64, etc.), more advanced features can be extracted.
[0127] Activation function: ReLU activation function (Rectified Linear Unit) is used to increase the nonlinear expression ability of the model;
[0128] Pooling layer: Use the Max Pooling layer to downsample the image feature map, reduce the number of parameters, and improve the stability of the features;
[0129] Feature extraction:
[0130] The convolution layer learns features, and each convolution kernel can be automatically optimized during the training process to extract features such as edges, textures, and shapes in the image. In the later stage of the model, a fully connected layer can be used to map the features to the output layer to achieve the final image classification or enhancement task;
[0131] S2. Generative Adversarial Network (GAN)
[0132] GAN Architecture:
[0133] Generator:
[0134] Input: A random noise vector in the latent space (e.g. a 100-dimensional vector).
[0135] Network structure: It consists of multiple fully connected layers and transposed convolutional layers to transform the latent vector into an enhanced night image;
[0136] Output: A network that generates close to real (enhanced) images;
[0137] Discriminator:
[0138] Input: real augmented images and images generated by the generator;
[0139] Network structure: Use convolutional layers to extract features, and finally use the Sigmoid activation function to output the true probability of the image (0 for false, 1 for true).
[0140] Training process:
[0141] GAN training is an adversarial process. The generator competes with the discriminator. The generator strives to improve the authenticity of the image, while the discriminator strives to distinguish between real and fake images. After many iterations, the generator can generate high-quality enhanced images with significantly improved detail restoration capabilities.
[0142] S3. Variational Autoencoder (VAE)
[0143] VAE Architecture:
[0144] Encoder:
[0145] Network structure: Using convolutional layers, the input nighttime image is encoded into distribution parameters (mean and variance) of the latent space;
[0146] Decoder:
[0147] Network structure: Use the transposed convolution layer to reconstruct the points in the latent space back to the space of the original image, with the goal of generating an enhanced image;
[0148] Training objectives:
[0149] VAE optimizes reconstruction loss and Kullback-Leibler divergence (KL divergence) during training, and can learn the potential representation of images, so that the model can not only reconstruct the input image, but also generate new high-quality images through the latent space;
[0150] S4. Comprehensive application
[0151] Combining CNN with GAN or VAE can form an efficient deep learning model for night image enhancement:
[0152] First, CNN is used to extract the features of night images;
[0153] Then, GAN or VAE is used to generate and enhance the image to improve details and brightness.
[0154] In the process of image enhancement, noise suppression is an important step, which can significantly improve the visual quality of the image and the reliability of subsequent processing. Specifically, it is Stirk filtering. The steps are as follows:
[0155] Fourier transform: Perform a Fourier transform on an image to convert the image from the spatial domain to the frequency domain.
[0156] Low-pass filter design: Design a Nyquist low-pass filter, select an appropriate cutoff frequency, and remove high-frequency components above this frequency.
[0157] Frequency domain multiplication: Multiply the low-pass filter by the Fourier transformed image element-wise.
[0158] Inverse transform: The result is converted back to the spatial domain through inverse Fourier transform to obtain the denoised image.
[0159] In the post-processing module of image processing, it is very important to adjust the brightness, contrast and color of the enhanced image to ensure that the final output image has a natural and beautiful visual effect. The specific steps are as follows:
[0160] S1. Adjust brightness
[0161] Here’s how:
[0162] Linear adjustment: Increase or decrease the brightness of all pixels by simply adding a constant value (brightness increment). For each pixel value I(x,y), the adjusted new value is:
[0163] I′(x,y)=I(x,y)+δ
[0164] Among them, \delta is the adjusted brightness value, and it is necessary to ensure that the new value does not exceed the valid range (for example, 0-255). Gamma transformation: Gamma transformation can adjust the brightness nonlinearly:
[0165] I'(x,y)=255\cdot\left(\frac{I(x,y)}{255}\right)^{\gamma}
[0166] Among them, \gamma is the gamma value, \gamma<increases brightness, \gamma> reduces brightness,
[0167] S2. Adjust contrast: Contrast adjustment is used to enhance the difference between light and dark areas in an image and improve the visual hierarchy. The specific methods are as follows:
[0168] Simple linear contrast adjustment:
[0169] I'(x,y)=\alpha\cdot(I(x,y)-128)+128
[0170] Among them, \alpha>1 enhances the contrast, and it is necessary to ensure that the new value is still within the valid range;
[0171] Histogram equalization: It can enhance the contrast by changing the gray value distribution of the image, which is especially suitable for low-contrast images.
[0172] Adjust color: Color adjustment includes adjusting saturation and hue to ensure that the color is natural and attractive. The method is as follows:
[0173] Saturation adjustment: You can increase or decrease the saturation by mixing the color value with the grayscale image. Saturation adjustment can specify a ratio:
[0174] I'(x,y)=I(x,y)+S\cdot(I(x,y)-I_{gray}(x,y))
[0175] Where I_{gray} is the gray value of the same pixel;
[0176] Hue adjustment: Change the overall hue of the image by rotating the hue ring and using the HSL model for processing;
[0177] Overall image adjustment steps:
[0178] Brightness adjustment: Improve the brightness of the image as needed;
[0179] Contrast enhancement: Enhance the contrast by linear transformation or histogram equalization;
[0180] Color correction: increase saturation and adjust hue to ensure natural colors;
[0181] Visual evaluation: Perform visual checks after processing to ensure that the image maintains a natural appearance for the user’s aesthetic experience;
[0182] Final output: After completing the above adjustments, you can use the image output function to save the image, output the final result in a common format, and ensure the quality of the output image. Through these steps, the post-processing module can effectively improve the visual effect of the image, ensuring that the enhanced image is not only clear and visible, but also visually natural and beautiful.
[0183] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A nighttime image enhancement and noise suppression integrated system based on deep learning, characterized by: System includes Image preprocessing module: normalizes the input nighttime images to improve the accuracy of subsequent processing; Deep learning model module: Using convolutional neural networks, we designed a deep learning model specific to nighttime image processing. The model can automatically learn image features and perform adaptive enhancement. Models that use generative adversarial networks or variational autoencoders to enhance image details and brightness; Noise Suppression Module: Combined with the output of the image enhancement process, the corresponding noise suppression algorithm is designed to effectively remove random noise in the image; Achieve adaptive processing of different noise types and improve the generalization ability of the system; Post-processing module: adjusts the brightness, contrast and color of the enhanced image to ensure the visual naturalness and beauty of the final output image.
2. The deep learning-based nighttime image enhancement and noise suppression integrated system according to claim 1, characterized in that: The image preprocessing module includes the following: S1. Normalization processing: Normalization is an important step in image preprocessing. Its main purpose is to convert the input image data into a uniform range, [0, 1] or [-1, 1]. Through normalization, the difference between images under different lighting conditions can be reduced, thereby improving the accuracy of subsequent processing. The following methods can be used for specific implementation: Linear Normalization: Linearly map the pixel values of the input image to a certain range. The formula is: Where I is the original pixel value, I_{\text{min}} and I_{\text{max}} are the minimum and maximum pixel values of the image, respectively, I' is the normalized pixel value, and Z-score standardization: the mean and standard deviation of each channel of the image are standardized to eliminate the differences in illumination and contrast between different images: I'=\frac{I-\mu}{\sigma} Among them, \mu is the mean of the image pixels, \sigma is the standard deviation; S2. Noise removal: Before normalizing the image, some simple filters can be used to reduce the noise in the image, which helps improve the effect of subsequent steps; Noise removal methods include the following: Mean filtering: average the pixels around each pixel in the image to reduce random noise; Median filtering: effectively removes salt and pepper noise by replacing the pixel value with the median of its neighborhood; Bilateral filtering: Considering the spatial distance between pixels and the similarity of pixel values, it can better retain edge information while removing noise; S3, Image enhancement: After normalization, some image enhancement techniques are applied to improve the visual effect of the image, especially in low-light environments; Enhancement methods include: Histogram equalization: By adjusting the histogram of the image, the pixel distribution of the image is made more uniform and the contrast is improved; Gamma transform: used to adjust the brightness of the image, increasing the brightness of the darker parts of the image through nonlinear transformation; Laplace filtering: Enhances the edge details of the image and improves the clarity of the image.
3. The deep learning-based nighttime image enhancement and noise suppression integrated system according to claim 1, characterized in that: Deep Learning Model Module In the deep learning model of night image processing, the selection of convolutional neural network combined with generative adversarial network or variational autoencoder to realize automatic learning of image features and adaptive enhancement has significant effects, including the following: S1. Convolutional Neural Network Network architecture design: Input layer: receives normalized night images with a size of 224x224 or other standard sizes; Convolutional layer: Use multiple convolutional layers to extract image features. The convolution kernel size can be 3x3 or 5x5. The number of convolution kernels is increased layer by layer to extract more advanced features. Activation function: ReLU activation function is used to increase the nonlinear expression ability of the model; Pooling layer: Use the maximum pooling layer to downsample the image feature map, reduce the number of parameters, and improve the stability of the features; Feature extraction: The convolution layer learns features, and each convolution kernel can be automatically optimized during the training process to extract features such as edges, textures, and shapes in the image. In the later stage of the model, a fully connected layer can be used to map the features to the output layer to achieve the final image classification or enhancement task; S2. Generative Adversarial Network GAN Architecture: Generator: Input: A random noise vector in the latent space. Network structure: It consists of multiple fully connected layers and transposed convolutional layers to transform the latent vector into an enhanced night image; Output: Generates an image close to the real one; Discriminator: Input: real augmented images and images generated by the generator; Network structure: Use convolutional layers to extract features, and finally use the Sigmoid activation function to output the true probability of the image. Training process: GAN training is an adversarial process. The generator competes with the discriminator. The generator strives to improve the authenticity of the image, while the discriminator strives to distinguish between real and fake images. After many iterations, the generator can generate high-quality enhanced images with significantly improved detail restoration capabilities. S3. Variational Autoencoder VAE Architecture: Encoder: Network structure: Use convolutional layers to encode the input nighttime image into distribution parameters in the latent space; Decoder: Network structure: Use the transposed convolution layer to reconstruct the points in the latent space back to the space of the original image, with the goal of generating an enhanced image; Training objectives: VAE optimizes reconstruction loss and Kullback-Leibler divergence during training, and can learn the potential representation of images, so that the model can not only reconstruct the input image, but also generate new high-quality images through the latent space; S4. Comprehensive application Combining CNN with GAN or VAE can form an efficient deep learning model for night image enhancement: First, CNN is used to extract the features of night images; Then, GAN or VAE is used to generate and enhance the image to improve details and brightness.
4. The deep learning-based nighttime image enhancement and noise suppression integrated system according to claim 1, characterized in that: In the process of image enhancement, noise suppression is an important step that can significantly improve the visual quality of the image and the reliability of subsequent processing. The steps are as follows: Fourier transform: Perform Fourier transform on the image to convert the image from the spatial domain to the frequency domain; Low-pass filter design: Design a Nyquist low-pass filter, select an appropriate cutoff frequency, and remove high-frequency components above this frequency; Frequency domain multiplication: multiply the low-pass filter by the Fourier transformed image element by element; Inverse transform: The result is converted back to the spatial domain through inverse Fourier transform to obtain the denoised image.
5. The deep learning-based nighttime image enhancement and noise suppression integrated system according to claim 1, characterized in that: In the post-processing module of image processing, it is very important to adjust the brightness, contrast and color of the enhanced image to ensure that the final output image has a natural and beautiful visual effect. The specific steps are as follows: S1. Adjust brightness Here’s how: Linear adjustment: Increase or decrease the brightness of all pixels by simply adding a constant value (brightness increment). For each pixel value I(x,y), the adjusted new value is: I′(x,y)=I(x,y)+δ Among them, \delta is the adjusted brightness value, and it is necessary to ensure that the new value does not exceed the valid range (for example, 0-255). Gamma transformation: Gamma transformation can adjust the brightness nonlinearly: I'(x,y)=255\cdot\left(\frac{I(x,y)}{255}\right)^{\gamma}where \gamma is the gamma value, \gamma<increases brightness, \gamma> reduces brightness, S2. Adjust contrast: Contrast adjustment is used to enhance the difference between light and dark areas in an image and improve the visual hierarchy. The specific methods are as follows: Simple linear contrast adjustment: I'(x,y)=\alpha\cdot(I(x,y)-128)+128 Among them, \alpha>1 enhances the contrast, and it is necessary to ensure that the new value is still within the valid range; Histogram equalization: It can enhance the contrast by changing the gray value distribution of the image, which is especially suitable for low-contrast images. Adjust color: Color adjustment includes adjusting saturation and hue to ensure that the color is natural and attractive. The method is as follows: Saturation adjustment: You can increase or decrease the saturation by mixing the color value with the grayscale image. Saturation adjustment can specify a ratio: I'(x,y)=I(x,y)+S\cdot(I(x,y)-I_{gray}(x,y)) Where I_{gray} is the gray value of the same pixel; Hue adjustment: Change the overall hue of the image by rotating the hue ring and using the HSL model for processing; Overall image adjustment steps: Brightness adjustment: Improve the brightness of the image as needed; Contrast enhancement: Enhance the contrast by linear transformation or histogram equalization; Color correction: increase saturation and adjust hue to ensure natural colors; Visual evaluation: Perform visual checks after processing to ensure that the image maintains a natural appearance for the user’s aesthetic experience; Final output: After completing the above adjustments, you can use the image output function to save the image, output the final result in a common format, and ensure the quality of the output image. Through these steps, the post-processing module can effectively improve the visual effect of the image, ensuring that the enhanced image is not only clear and visible, but also visually natural and beautiful.
Citation Information
Cited By
Organ-like cell microcarrier detection method and organ-like cell microcarrier model building method
CN119151909A
Visual promotion method for light contrast inside and outside large ship unloading cabin
CN119251056A
Endoscopic pipeline inner wall crack automatic identification system based on video deep learning
CN119693841A
Driver fatigue driving state identification method based on deep learning
CN119723537A
Image low-light enhancement joint deblurring method based on bilevel planning light and shadow clearness network
CN119831878A
