Deep learning-based denoising and enhancement system and method for CMOS cameras

Abstract

The present invention discloses a denoising and enhancement system and method for a CMOS camera based on deep learning. The denoising and enhancement system includes: an acquisition unit, which collects paired low-light images and corresponding reference images under normal illumination as a training data set by building a low-light environment simulation and acquisition system; a training unit, which pre-processes the data set obtained by the acquisition unit and inputs it into a convolutional neural network in pairs, and trains the model by optimizing the loss function; a testing unit, which inputs the image to be tested into the trained model to obtain a denoised and enhanced image, and then performs other image processing operations. The present invention denoises and enhances the pictures taken by the CMOS camera in a dark environment with only weak visible light and infrared light, and solves problems such as the need to synthesize noise, perform noise estimation, incomplete noise removal, or excessive smoothing of image details in previous denoising algorithms.

Description

Technical Field

[0001] The present invention relates to a denoising and enhancement system and method, and in particular to a denoising and enhancement system and method for a CMOS camera based on deep learning. Background Art

[0002] Under conditions of sufficient light, the images captured by the image acquisition device have a high dynamic range and rich scene detail information. In an environment with severe lack of light, there is usually only weak visible light and some infrared light. Compared with ordinary cameras, low-light CMOS cameras can be used to image in darker and more extreme environments, as well as under normal light. This is of great significance in the fields of night vision, security monitoring, drones, military, field surveys, etc. However, in an environment with extreme lack of light, the imaging of low-light CMOS cameras still faces the problem of low signal-to-noise ratio, and the noise will drown out the useful information of the image.

[0003] In recent years, there have been many studies on image denoising, but previous studies usually simplify the denoising problem into two parts: the original clear image plus additive noise. Denoising is achieved by subtracting additive noise from the noise-contaminated image, and the noise is usually assumed to be a Gaussian white noise model. The image noise component obtained by CMOS sensors is not only related to the circuit and camera pipeline, but also to the light. As the light decreases, the noise will become more and more obvious, and the noise component will become more complex. A single denoising algorithm cannot fully meet the denoising needs of low-light CMOS cameras. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a denoising and enhancement system and method for a CMOS camera based on deep learning.

[0005] The present invention solves the above technical problems through the following technical solutions: a denoising and enhancement system for a CMOS camera based on deep learning, characterized in that it comprises:

[0006] An acquisition unit, by building a low-light environment simulation and acquisition system, collects paired low-light images and corresponding reference images under normal illumination as a training data set; the acquisition unit includes a camera module, a shooting control module, a shooting object module, an illumination module, and an illuminance measurement module. The camera module includes a CMOS camera and an optical platform. The CMOS camera is fixed on the optical platform. The camera module is used for shooting object imaging and controlling optical parameters such as aperture, exposure, and focal length to ensure clear images. The shooting control module includes a computer, which is connected to the CMOS camera in the camera module and displays the images captured by the CMOS camera. The shooting control module is used for controlling the shooting time, saving images, and cooperating with the camera module to preview images. The shooting object module obtains the training data set by replacing the objects to be shot with different colors and reflectivities. The illumination module includes an adjustable visible light source, an infrared supplementary light source, and a laboratory illumination light source. The illumination module is used for laboratory illumination and simulating the light source spectrum and brightness in the actual outdoors. The illuminance measurement module measures the illuminance of the experimental light source and cooperates with the illumination module to control the illuminance of the experimental light source;

[0007] A training unit, by preprocessing the data set obtained by the acquisition unit and inputting it into a convolutional neural network in pairs, and training the model by optimizing the loss function; the training unit includes an image preprocessing module, an encoding module, a decoding module, and an optimization module. The image preprocessing module performs translation cropping and flipping data augmentation operations on the obtained original images in pairs. The encoding module extracts image features and reduces the amount of calculation by performing convolution and activation operations on the images. The decoding module restores the image size by performing deconvolution, activation, and residual connection operations on the images. The optimization module trains to improve the accuracy and robustness of the model by reducing the loss function;

[0008] A testing unit, by inputting the image to be tested into the trained model, obtains the denoised and enhanced image, and then performs other image processing operations; the testing unit includes a test denoising and enhancement module and an image post-processing module. The test denoising and enhancement module inputs the new image to be tested into the trained model and calculates to obtain the denoised and enhanced image. The image post-processing module further improves the image quality according to the requirements by connecting operations such as image white balance, brightness, and contrast adjustment.

[0009] Preferably, the encoding module performs forward propagation calculation; the deconvolution of the decoding module restores the image size through the calculation method of deconvolution, and the residual connection directly merges the operation results of the corresponding layers of the encoding module and the feature maps of the corresponding layers of the decoding module on the channels.

[0010] Preferably, the test denoising and enhancement module takes inputs in batches, and directly outputs the denoised and enhanced images through GPU parallel computing by calling the trained model and parameter configuration file in the training unit; the white balance of the image post-processing module adopts the perfect reflection algorithm, and the brightness adjustment controls the over-dark or over-exposed images through histogram equalization and Gamma correction.

[0011] Preferably, the adjustable visible light source and the infrared supplementary light source are used to simulate the light source spectrum and brightness in the actual outdoor environment.

[0012] Preferably, the laboratory lighting source is located at the top of a laboratory, and the scene is transformed into a normal lighting environment and a low-illuminance environment through a switch.

[0013] Preferably, the encoding module includes five convolutional units, and the decoding module includes four convolutional units.

[0014] The present invention also provides a denoising and enhancement method for a CMOS camera based on deep learning, which is characterized by comprising the following steps:

[0015] Step 1, fix the camera module and set the shooting control of the shooting control module; the camera module includes a CMOS camera and an optical platform, the CMOS camera is fixed on the optical platform, and the height and direction remain unchanged all the time; the shooting control module includes a computer, the computer is connected to the CMOS camera in the camera module, displays the images captured by the CMOS camera, and sets shooting parameters such as resolution, time-lapse shooting, storage name and address.

[0016] Step 2, adjust the object to be photographed and the CMOS camera so that the object to be photographed is within the shooting field of view of the low-light CMOS camera, and adjust the optical parameters to ensure clear imaging.

[0017] Step 3, set the adjustable visible light source and the infrared supplementary light source in the lighting module to provide uniform illumination for the object to be photographed.

[0018] Step 4, the illuminance measurement module measures the illuminance of the experimental light source and cooperates with the lighting module to control the illuminance of the experimental light source.

[0019] Step 5, transform the scene into a normal lighting environment and a low-illuminance environment by switching the laboratory lighting source 7 in the lighting module, and shoot paired low-illuminance images and normal lighting reference images as the training data set.

[0020] Step 6, the shooting object module improves the richness of the data set by replacing the objects to be photographed with different colors and reflectivities, and quickly and massively obtains the training data set to complete the shooting and data set construction.

[0021] Step 7: Construct a convolutional neural network, which includes an encoding module and a decoding module, and import the training data set for training; the optimization module updates the weight parameter configuration by reducing the loss function to train and improve the accuracy and robustness of the model.

[0022] Step 8: The test denoising and enhancement module inputs the image to be denoised into the trained denoising network, and the output image is the denoised and enhanced image.

[0023] Step 9: Perform image post-processing operations such as image white balance, brightness, and contrast adjustment through the image post-processing module.

[0024] Preferably, the encoding module performs forward propagation by performing convolution and activation operations on the image to complete the extraction of image features and reduce the computational amount.

[0025] Preferably, the decoding module performs feature upsampling to restore the image size by performing deconvolution, activation, and residual connection operations on the image.

[0026] Preferably, the test denoising and enhancement module inputs the new image in batches to increase the image processing speed, and directly outputs the denoised and enhanced image through GPU parallel computing by calling the trained model and parameter configuration file in the training unit.

[0027] The positive and progressive effects of the present invention are as follows: The present invention constructs a low-light environment simulation and acquisition system and collects paired low-light images and corresponding reference images under normal light as the training data set. By constructing a convolutional neural network as the denoising network, the obtained training data is integrated and imported for training. Finally, the image to be denoised is input into the trained denoising network, and the denoised and enhanced image is output. According to the requirements of the application scenario, the present invention denoises and enhances the images taken by the CMOS camera in a dark environment containing only weak visible light and infrared light, solving problems such as the need to synthesize noise, perform noise estimation, incomplete noise removal, or excessive smoothing of image details in previous denoising algorithms. Description of the Drawings

[0028] Figure 1 It is the principle block diagram of the denoising and enhancement system of the CMOS camera based on deep learning of the present invention.

[0029] Figure 2 It is the working principle diagram of the present invention.

[0030] Figure 3 It is the flowchart of the denoising and enhancement method of the CMOS camera based on deep learning of the present invention. Detailed Embodiments

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0032] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.

[0033] Traditional methods and the latest denoising methods in existing research usually focus on the removal of Gaussian noise, and obtain a large number of noisy images by adding noise to clean images in batches, or artificially increase and decrease the exposure time to obtain normal illumination images and low-illumination images. However, both the low-illumination images and normal illumination reference images of the present invention are all collected from the experimental device. The features learned by using the deep learning method are specific to the training data and more in line with the requirements of the application scenario. In addition, the advantage of deep learning is its strong generalization ability, fast parallel computing speed on the GPU, and the end-to-end feature avoids cumbersome intermediate image processing steps. Applying the convolutional neural network to the denoising and enhancement of low-light CMOS camera images, a large number of stable, reliable and uniquely noisy training data sets can be quickly obtained through the established low-illumination environment simulation and acquisition system, and good denoising effects can be achieved, solving the problems that the previous denoising algorithms need to perform noise estimation, the noise removal of low-light CMOS cameras is not clean, or the images are over-smoothed.

[0034] As Figure 1 and Figure 2 shown, the denoising and enhancement system of the CMOS camera based on deep learning of the present invention includes:

[0035] An acquisition unit, by building a low-light environment simulation and acquisition system, collects paired low-light images and corresponding reference images under normal illumination as a training data set; the acquisition unit includes a camera module, a shooting control module, a shooting object module, an illumination module, and an illuminance measurement module. The camera module includes a CMOS camera 1 and an optical platform 2. The CMOS camera 1 is fixed on the optical platform 2. The camera module is used to capture object images and control optical parameters such as aperture, exposure, and focal length to ensure clear images. The shooting control module includes a computer 3. The computer 3 is connected to the CMOS camera 1 in the camera module to display the images captured by the CMOS camera 1. The shooting control module is used to control the shooting time, save images, and cooperate with the camera module to preview images. The shooting object module obtains a training data set by replacing the objects to be photographed 8 with different colors and reflectivities. The illumination module includes an adjustable visible light source 5, an infrared supplementary light source 6, and a laboratory illumination light source 7. The illumination module is used for laboratory illumination and simulating the light source spectrum and brightness of the actual outdoors. The illuminance measurement module measures the illuminance of the experimental light source and cooperates with the illumination module to control the illuminance of the experimental light source.

[0036] A training unit, by preprocessing the data set obtained by the acquisition unit and inputting it into a convolutional neural network in pairs, trains the model by optimizing the loss function; the training unit includes an image preprocessing module, an encoding module, a decoding module, and an optimization module. The image preprocessing module performs data augmentation operations such as translation cropping and flipping on the obtained original images in pairs. The encoding module extracts image features and reduces the amount of calculation by performing convolution and activation operations on the images. The decoding module performs feature upsampling to restore the image size by performing deconvolution, activation, and residual connection operations on the images. The optimization module trains to improve the accuracy and robustness of the model by reducing the loss function.

[0037] A testing unit, by inputting the image to be tested into the trained model, obtains a denoised and enhanced image, and then performs other image processing operations; the testing unit includes a test denoising and enhancement module and an image post-processing module. The test denoising and enhancement module inputs the new image to be tested into the trained model and calculates to obtain a denoised and enhanced image. The image post-processing module further improves the image quality according to requirements by connecting operations such as image white balance, brightness, and contrast adjustment.

[0038] The image preprocessing module can cut out image blocks with a size of 512×512 at different positions in the original image, and flip the 512×512 image blocks by 90°, 180°, and 270° to complete the data augmentation operation and enrich the data set.

[0039] The encoding module performs forward propagation calculations, using a convolutional kernel with a size of 3×3, a stride of 2, and a "same" padding method for convolution operations. The result of the operation is subjected to max pooling downsampling with a size of 2×2, and the activation function selects ReLU (Rectified Linear Unit). The transposed convolution of the decoding module restores the image size through the calculation method of transposed convolution. The residual connection directly merges the operation result of the corresponding layer of the encoding module and the feature map of the corresponding layer of the decoding module on the channel.

[0040] The test denoising and enhancement module is input in batches to increase the image processing speed. By calling the trained model and parameter configuration file in the training unit, the GPU (Graphics Processing Unit) performs parallel computing processing to directly output the denoised and enhanced image. The white balance of the image post-processing module adopts the perfect reflection algorithm, and the brightness adjustment controls the over-dark or over-exposed image through histogram equalization method and Gamma correction.

[0041] The object 8 to be photographed can be different objects with different colors and reflectivities, having rich diversity and differences, covering the characteristics of common objects. The distance between the object 8 and the CMOS camera 1 is between 0.71 and 0.73 m. Adjust the focal length and aperture of the CMOS camera 1 to ensure clear imaging of the image displayed on the computer 3.

[0042] The adjustable visible light source 5 and the infrared supplementary light source 6 are used to simulate the light source spectrum and brightness in the actual outdoor environment. They are located on the same side of the CMOS camera 1 to provide uniform illumination for the object 8 to be photographed. The illuminance measurement module includes a micro-illuminance meter 4 for convenient use and real-time measurement. One adjustable visible light source 5 and one infrared supplementary light source 6 are located on the same side of the CMOS camera 1, and the probe of a micro-illuminance meter 4 is placed beside the object 8 to be photographed. Multiple laboratory lighting sources 7 are located on the top of the laboratory to control the lighting. Since the CMOS camera 1 is fixed on the optical platform 2, the jitter of the CMOS camera 1 and the influence of the environment are avoided, and pictures with the same size and the same position are obtained under different lighting conditions. By replacing the object 8 to be photographed with different colors and reflectivities, a large number of paired low-illuminance images and corresponding reference images under normal lighting can be quickly obtained as the training data set.

[0043] The CMOS camera 1 is controlled by connecting to a computer 3. The computer 3 sets the parameters before shooting and performs a time-lapse shooting. It remains closed during shooting to avoid its light emission causing changes in illumination. The object to be photographed is within the shooting field of view of the CMOS camera 1, and the focal length is adjusted to keep the captured image clear. The adjustable visible light source 5 and the infrared supplementary light source 6 are used to simulate the light source spectrum and brightness in the actual outdoors, providing uniform illumination for the object to be photographed 8. The main body of the micro-illuminance meter 4 is located inside a box to avoid its light causing changes in illumination. The illuminance value is read inside the box. The probe of the micro-illuminance meter 4 is placed beside the object to be photographed to measure the illumination level around the object to be photographed 8. Together with the adjustable visible light source 5, it controls the brightness at an extremely low level and quantifies the illumination level felt by the human eye. The reading of the micro-illuminance meter 4 is about 0.002 lux. The laboratory lighting source 7 is located on the top of a laboratory, and the scene is transformed into a normal illumination environment and a low-illumination environment through a switch.

[0044] The convolutional neural network is part of the training unit and consists of two modules: an encoding module and a decoding module. Its functions are to extract features, increase the receptive field of the image, and reduce the amount of computation. The encoding module contains five convolutional units:

[0045] The first convolutional unit includes a 3×3 convolutional layer and ReLU activation, followed by a second 3×3 convolution and ReLU activation, and then max-pooling downsampling.

[0046] The second convolutional unit includes a 3×3 convolutional layer and ReLU activation, followed by a second 3×3 convolution and ReLU activation, and then max-pooling downsampling.

[0047] The third convolutional unit includes a 3×3 convolutional layer and ReLU activation, followed by a second 3×3 convolution and ReLU activation, and then max-pooling downsampling.

[0048] The fourth convolutional unit includes a 3×3 convolutional layer and ReLU activation, followed by a second 3×3 convolution and ReLU activation, then Dropout, randomly discarding some parameters with the parameter value set to 0.5, and finally max-pooling downsampling.

[0049] The fifth convolutional unit includes a 3×3 convolutional layer and ReLU activation, followed by a second 3×3 convolution and ReLU activation, and finally randomly discarding some parameters with the function value set to 0.5.

[0050] The decoding module contains four convolutional units:

[0051] The sixth convolutional unit includes 2×2 transposed convolution upsampling, and merges with the feature map output by the fourth convolutional unit through a residual connection, followed by a 3×3 convolutional layer and ReLU activation, and then a second 3×3 convolution and ReLU activation.

[0052] The seventh convolutional unit includes a 2×2 transposed convolution for upsampling, and merges with the feature map output by the third convolutional unit through a residual connection, followed by a 3×3 convolutional layer and ReLU activation, and then a second 3×3 convolution and ReLU activation.

[0053] The eighth convolutional unit includes a 2×2 transposed convolution for upsampling, and merges with the feature map output by the second convolutional unit through a residual connection, followed by a 3×3 convolutional layer and ReLU activation, and then a second 3×3 convolution and ReLU activation.

[0054] The ninth convolutional unit includes a 2×2 transposed convolution for upsampling, and merges with the feature map output by the first convolutional unit through a residual connection, followed by a 3×3 convolutional layer and ReLU activation, and then a second 3×3 convolution and ReLU activation.

[0055] The convolutional neural network uses the loss function L1 to calculate the mean squared error MSE, as shown in the following equation (1):

[0056]

[0057] where I GT is the reference image, I Pre is the predicted image, i represents the horizontal pixel value of the image, j represents the vertical pixel value of the image, M is the width of the image, and N is the height of the image. The batch size is set to 1, the ratio of the training dataset to the validation set is set to 9:1, and a total of 2000 epochs (an epoch is a general chart library for application developers and visualization designers) are trained. The learning rate for the first 1000 epochs is set to 0.0001, and the learning rate for the last 1000 epochs is set to 0.0001. The Adam optimizer is selected, with parameters β1 = 0.9, β2 = 0.999, epsilon = 1×10 -8 , and β1, β2, and epsilon are all parameters of the Adam optimizer.

[0058] Select images from the test set and input them into the trained network to output the predicted denoised images, and perform image post-processing operations such as image white balance, brightness, and contrast adjustment, and finally complete the denoising and enhancement task of the low-light CMOS camera.

[0059] As Figure 3 shown, the denoising and enhancement method for a CMOS camera based on deep learning according to the present invention includes the following steps:

[0060] Step 1: Fix the camera module and set the shooting control of the shooting control module; the camera module includes a CMOS camera 1 and an optical platform 2. The CMOS camera 1 is fixed on the optical platform 2, keeping the height and direction unchanged all the time; the shooting control module includes a computer 3. The computer 3 is connected to the CMOS camera 1 in the camera module, displays the images captured by the CMOS camera 1, and sets shooting parameters such as resolution, time-lapse shooting, storage name, and address.

[0061] Step 2: Adjust the object to be photographed 8 and the CMOS camera 1 so that the object to be photographed 8 is within the shooting field of view of the low-light CMOS camera 1, and adjust the optical parameters to ensure clear imaging.

[0062] Step 3: Set the adjustable visible light source and the infrared supplementary light source in the lighting module to provide uniform illumination for the object to be photographed.

[0063] Step 4: The illuminance measurement module measures the light source illuminance of the experiment and cooperates with the lighting module to control the light source illuminance of the experiment.

[0064] Step 5: Change the scene to a normal lighting environment and a low-illuminance environment by turning on the laboratory lighting source 7 in the lighting module, and take paired low-illuminance images and normal lighting reference images as the training data set.

[0065] Step 6: The object module to be photographed improves the richness of the data set by replacing the objects to be photographed 8 with different colors and reflectivities, and quickly obtains a large amount of training data sets in large quantities to complete shooting and data set construction.

[0066] Step 7: Construct a convolutional neural network. The convolutional neural network includes an encoding module and a decoding module, and import the training data set for training; the optimization module updates the weight parameter configuration by reducing the loss function to train and improve the accuracy and robustness of the model.

[0067] Step 8: The test denoising and enhancement module inputs the image to be denoised into the trained denoising network, and the output image is the denoised and enhanced image.

[0068] Step 9: Perform image post-processing operations such as image white balance, brightness, and contrast adjustment through the image post-processing module.

[0069] The encoding module performs forward propagation by performing convolution and activation operations on the image to complete image feature extraction and reduce the amount of calculation. The decoding module performs feature upsampling to restore the image size by performing deconvolution, activation, and residual connection operations on the image. The test denoising and enhancement module inputs the new image in batches to increase the image processing speed, and directly outputs the denoised and enhanced image by calling the trained model and parameter configuration file in the training unit for GPU parallel computing.

[0070] The present invention constructs a low - illumination environment simulation and acquisition system, collects paired low - illumination images and corresponding reference images under normal illumination as a training data set, constructs a convolutional neural network as a denoising network, integrates and imports the obtained training data pairs for training, and finally inputs the image to be denoised into the trained denoising network to output a denoised and enhanced image. According to the requirements of the application scenario, the present invention denoises and enhances the pictures taken by a CMOS camera in a dark environment containing only weak visible light and infrared light, solving problems such as the need to synthesize noise, perform noise estimation, incomplete noise removal, or excessive smoothing of image details in previous denoising algorithms.

[0071] The present invention can also combine other image - processing methods according to actual needs, such as image demosaicking, image super - resolution, de - fogging, de - raining and other algorithms to further improve the image quality. The embodiments described above are only descriptions of the implementation manners of the present invention and do not limit the scope of the present invention. Those skilled in the art can make various improvements and deformations to the present invention without departing from the design spirit of the present invention, but these improvements and deformations should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A denoising and enhancement system for a CMOS camera based on deep learning, characterized in that, It includes: An acquisition unit, which collects paired low-light images and corresponding reference images under normal illumination as a training data set by building a low-light environment simulation and acquisition system; The acquisition unit includes a camera module, a shooting control module, a shooting object module, an illumination module, and an illuminance measurement module. The camera module includes a CMOS camera and an optical platform. The CMOS camera is fixed on the optical platform. The camera module is used to capture object images and control optical parameters such as aperture, exposure, and focal length to ensure clear images. The shooting control module includes a computer, which is connected to the CMOS camera in the camera module and displays the images captured by the CMOS camera. The shooting control module is used to control the shooting time, save images, and cooperate with the camera module to preview images. The shooting object module obtains a training data set by replacing the objects to be photographed with different colors and reflectivities. The illumination module includes an adjustable visible light source, an infrared supplementary light source, and a laboratory illumination light source. The illumination module is used for laboratory illumination and simulating the light source spectrum and brightness of the actual outdoors. The illuminance measurement module measures the illuminance of the experimental light source and cooperates with the illumination module to control the illuminance of the experimental light source; A training unit, which pairs the data set obtained by the acquisition unit after image preprocessing and inputs it into a convolutional neural network, and trains the model by optimizing the loss function; the training unit includes an image preprocessing module, an encoding module, a decoding module, and an optimization module. The image preprocessing module performs translational cropping and flipping data augmentation operations on the obtained original images in pairs; the encoding module extracts image features and reduces the amount of calculation by performing convolution and activation operations on the images; The decoding module restores the image size by performing deconvolution, activation, and residual connection operations on the images; the optimization module trains to improve the accuracy and robustness of the model by reducing the loss function; A testing unit, which inputs the image to be tested into the trained model to obtain a denoised and enhanced image, and then performs other image processing operations; the testing unit includes a test denoising and enhancement module and an image post-processing module. The test denoising and enhancement module inputs the new image to be tested into the trained model and calculates to obtain a denoised and enhanced image; the image post-processing module further improves the image quality according to requirements by connecting operations such as image white balance, brightness, and contrast adjustment.

2. The denoising and enhancement system of the CMOS camera based on deep learning according to claim 1, characterized in that, The encoding module performs forward propagation calculations; the deconvolution of the decoding module restores the image size through the calculation method of deconvolution, and the residual connection directly merges the operation results of the corresponding layers of the encoding module and the feature maps of the corresponding layers of the decoding module on the channels.

3. The denoising and enhancement system of the CMOS camera based on deep learning according to claim 1, characterized in that, The test denoising and enhancement module is input in batches. By calling the trained model and parameter configuration file in the training unit, GPU parallel computing processes directly outputs the denoised and enhanced image; the white balance of the image post-processing module adopts the perfect reflection algorithm, and the brightness adjustment controls the image from being too dark or too bright through histogram equalization and Gamma correction.

4. The denoising and enhancement system of the CMOS camera based on deep learning according to claim 1, wherein, The adjustable visible light source and the infrared supplementary light source are used to simulate the light source spectrum and brightness of the actual outdoors.

5. The denoising and enhancement system of the CMOS camera based on deep learning according to claim 1, characterized in that The laboratory lighting source is located at the top of a laboratory, and the scene is transformed into a normal lighting environment and a low-illuminance environment through a switch.

6. The denoising and enhancement system of the CMOS camera based on deep learning according to claim 1, characterized in that, The encoding module contains five convolutional units, and the decoding module contains four convolutional units.

7. A denoising and enhancement method for a CMOS camera based on deep learning, characterized in that, It includes the following steps: Step 1, fix the camera module and set the shooting control of the shooting control module; the camera module includes a CMOS camera and an optical platform. The CMOS camera is fixed on the optical platform, and its height and direction remain unchanged all the time; the shooting control module includes a computer, which is connected to the CMOS camera in the camera module, displays the images captured by the CMOS camera, and sets shooting parameters such as resolution, delayed shooting, storage name, and address. Step 2, adjust the object to be photographed and the CMOS camera so that the object to be photographed is within the shooting field of view of the low-light CMOS camera, and adjust the optical parameters to ensure clear imaging. Step 3, set the adjustable visible light source and the infrared supplementary light source in the lighting module to provide uniform illumination for the object to be photographed. Step 4, the illuminance measurement module measures the illuminance of the experimental light source and cooperates with the lighting module to control the illuminance of the experimental light source. Step 5, transform the scene into a normal lighting environment and a low-illuminance environment by switching the laboratory lighting source in the lighting module, and take paired low-illuminance images and normal lighting reference images as the training data set. Step 6, the shooting object module improves the richness of the data set by replacing the objects to be photographed with different colors and reflectivities, and quickly obtains a large amount of training data sets to complete shooting and data set construction. Step 7, construct a convolutional neural network, which contains an encoding module and a decoding module, and import the training data set for training; the optimization module updates the weight parameter configuration by reducing the loss function to train and improve the accuracy and robustness of the model. Step 8, the test denoising and enhancement module inputs the image to be denoised into the trained denoising network, and the output image is the denoised and enhanced image. Step 9, perform image post-processing operations such as image white balance, brightness, and contrast adjustment through the image post-processing module.

8. The denoising and enhancement method of the CMOS camera based on deep learning according to claim 7, characterized in that, The encoding module performs forward propagation through convolutional and activation operations on the image to complete feature extraction of the image and reduce the amount of calculation.

9. The denoising and enhancement method of the CMOS camera based on deep learning according to claim 7, characterized in that, The decoding module performs feature upsampling to restore the image size through deconvolution, activation, and residual connection operations on the image.

10. The denoising and enhancement method of the CMOS camera based on deep learning according to claim 7, wherein The test denoising and enhancement module inputs new images in batches to increase the image processing speed, and directly outputs the denoised and enhanced images through GPU parallel computing by calling the trained model and parameter configuration file in the training unit.

Images