Denoising and enhancement processing method and system for sensor image data

Abstract

The present application relates to the field of image processing technology, specifically to a method and system for denoising and enhancing sensor image data. The method includes: preprocessing the original image data, detecting and classifying the noise type of the preprocessed image data through a deep learning model; constructing a high-fidelity denoising and enhancement strategy, including noise domain decomposition, dynamic area weighted denoising, and multi-scale enhancement and texture compensation; constructing an adaptive multi-domain collaborative denoising and enhancement strategy, combining noise domain adaptive projection based on variational autoencoders, multi-scale detail enhancement under a pyramid structure, and cross-domain texture migration and fusion; selecting the optimal method to denoise and enhance the image through the mapping relationship between noise detection results and enhancement strategies. The method of the present application can achieve high-fidelity denoising, detail restoration, and quality improvement of images in complex scenes, providing a reliable solution for sensor image data processing.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and specifically to a method and system for denoising and enhancing sensor image data. Background Art

[0002] The image data collected by the sensor is often affected by various noises, such as Gaussian noise, salt and pepper noise, Poisson noise, etc., which seriously affects the image quality and the effect of subsequent processing.

[0003] Traditional image denoising methods are typically designed for specific types of noise and lack adaptability to complex noise environments. While removing noise, these methods often oversmoothe image details, leading to texture distortion, blurred edges, and other issues. Furthermore, traditional methods are also limited in their effectiveness for image enhancement, making them difficult to meet the demands of high-quality image restoration and enhancement.

[0004] In recent years, deep learning-based image denoising and enhancement methods have garnered widespread attention. These methods achieve image denoising and enhancement by building deep neural network models to learn the mapping relationship between noise and image content from large amounts of data. However, existing deep learning methods are mostly designed for specific types of noise and scenes, with limited generalization capabilities and difficulty adapting to complex and changing noise environments. Furthermore, these methods often neglect the preservation and enhancement of image texture details while denoising, resulting in limited improvement in denoised image quality.

[0005] To sum up, how to develop an adaptive, efficient, high-quality sensor image denoising and enhancement method for complex noise environments, make full use of domain knowledge, and improve the image restoration and enhancement effects is a key technical problem that needs to be solved urgently.

[0006] In view of this, the present application proposes a method and system for denoising and enhancing sensor image data. Summary of the Invention

[0007] To achieve the above objectives, this application provides a method and system for denoising and enhancing sensor image data. The specific technical solutions are as follows:

[0008] The denoising and enhancement processing method of sensor image data includes:

[0009] Acquire image data collected by the sensor and preprocess the image data;

[0010] Build a deep learning model to detect the noise type of preprocessed image data and determine the image denoising and enhancement strategy based on the noise type;

[0011] Build high-fidelity denoising and enhancement strategies for complex noise scenarios, including noise domain decomposition, dynamic region weighted denoising, and multi-scale enhancement and texture compensation;

[0012] Constructing adaptive multi-domain collaborative denoising and enhancement strategies, including noise domain adaptive projection based on variational autoencoder models, multi-scale detail enhancement under pyramid structures, and cross-domain texture transfer and fusion;

[0013] According to noise detection and classification, the corresponding denoising and enhancement strategies are selected, the sensor image is denoised and enhanced, and the processed sensor image is output.

[0014] Preferably, image data collected by the sensor is obtained, and grayscale conversion, normalization, image data smoothing, image data sharpening, and image data quantization are performed on the obtained image data; after the image data is pre-processed, it is used for noise type detection;

[0015] Construct a convolutional neural network (CNN) model for noise detection and classification. The model input is preprocessed noise image data, and the output is the probability distribution of noise types.

[0016] The convolutional neural network (CNN) model includes an input layer, a convolution layer, a pooling layer, and a fully connected layer. The CNN model is trained using an image dataset with noise type labels, a cross entropy loss function, and an Adam optimization algorithm.

[0017] The trained CNN model is used to classify noise. The preprocessed noise image data is input into the trained CNN model to obtain the probability distribution of noise types, and the category with the highest probability is selected as the final noise type.

[0018] Preferably, the noise domain decomposition includes: performing wavelet transform on the noisy image to decompose the image into low-frequency sub-bands and high-frequency sub-bands, performing adaptive threshold denoising on the high-frequency sub-bands, and adaptively adjusting the threshold according to the noise standard deviation; performing inverse wavelet transform on the low-frequency sub-bands and the denoised high-frequency sub-bands to reconstruct the denoised image.

[0019] Preferably, the dynamic area weighted denoising includes: dividing the denoised image into overlapping blocks; for each overlapping block, searching for similar overlapping blocks in the neighborhood to construct a similar block set, and determining similar overlapping blocks using Euclidean distance; performing collaborative denoising on each similar block set to obtain denoised overlapping blocks; and performing weighted averaging on the denoised overlapping blocks to obtain the final denoised image.

[0020] Preferably, the multi-scale enhancement and texture compensation include: performing multi-scale decomposition on the final denoised image to obtain sub-band coefficients at different scales; enhancing the sub-band coefficients of each scale to obtain enhanced sub-band coefficients; based on the multi-scale enhancement, performing texture compensation on the high-frequency sub-band to restore the texture details of the image; and performing multi-scale reconstruction on the enhanced sub-band coefficients and the texture-compensated high-frequency sub-band to obtain an enhanced image.

[0021] Preferably, the variational autoencoder model includes: a method for mapping a noise image to a latent space to achieve adaptive projection in the noise domain; the variational autoencoder model consists of an encoder and a decoder; the encoder maps the noise image to the posterior distribution of the latent variable; the decoder generates a reconstructed image from the latent variable; the objective function of the variational autoencoder model includes a reconstruction loss and a KL divergence regularization term, and the variational autoencoder model is trained by the objective function; using the trained variational autoencoder model, the noise image is mapped to the latent space to obtain an adaptive noise representation; and based on the noise representation, a corresponding denoised image is generated.

[0022] Preferably, the multi-scale detail enhancement under the pyramid structure includes: performing Gaussian pyramid decomposition on the denoised image to obtain image pyramids at different scales; constructing a Laplacian pyramid based on the image pyramids at different scales to extract multi-scale detail information of the image; adaptively enhancing each layer of the Laplacian pyramid to obtain an enhanced Laplacian pyramid; and reconstructing the enhanced Laplacian pyramid to obtain a detail-enhanced image.

[0023] Preferably, the cross-domain texture migration and fusion includes: extracting reference texture features from a reference image to guide the texture migration of the noisy image; using the reference texture features to perform texture migration on the detail-enhanced image to generate a texture-enhanced image; and performing multi-domain fusion of the texture-enhanced image and the original denoised image to obtain the final enhanced image.

[0024] Preferably, for a given noise image, a probability distribution of noise types is obtained through a trained convolutional neural network (CNN) model, and the noise type with the highest probability is selected according to the probability distribution;

[0025] Based on the noise detection and classification results, a mapping relationship between noise type and denoising and enhancement strategy is established; according to the mapping relationship, the corresponding denoising and enhancement strategy is selected, the noisy image is denoised and enhanced, and the corresponding processing results are output.

[0026] A denoising and enhancement processing system for sensor image data, comprising: a data acquisition module, a noise detection module, a first denoising and enhancement module, a second denoising and enhancement module, and an image output module;

[0027] The data acquisition module is used to acquire image data collected by the sensor and pre-process the image data;

[0028] The noise detection module is used to build a deep learning model, detect the noise type of the preprocessed image data, and determine the image denoising and enhancement strategy according to the noise type;

[0029] The first denoising and enhancement module is used to construct a high-fidelity denoising and enhancement strategy in complex noise scenes, including noise domain decomposition, dynamic area weighted denoising, and multi-scale enhancement and texture compensation;

[0030] The second denoising and enhancement module is used to construct an adaptive multi-domain collaborative denoising and enhancement strategy, including noise domain adaptive projection based on a variational autoencoder model, multi-scale detail enhancement under a pyramid structure, and cross-domain texture migration and fusion;

[0031] The image output module selects corresponding denoising and enhancement strategies according to noise detection and classification, performs denoising and enhancement processing on the sensor image, and outputs the processed sensor image.

[0032] Beneficial effects of this application: This application improves the effects of subsequent noise detection and denoising enhancement by preprocessing the image data collected by the sensor.

[0033] This application uses a deep learning model to adaptively detect noise types and select the optimal denoising and enhancement strategies for different noise types, thereby improving the adaptability and robustness of the method.

[0034] This application constructs a high-fidelity denoising and enhancement strategy for complex noise scenes, effectively suppressing noise while protecting image details, and improving the quality of denoising and enhancement.

[0035] This application introduces an adaptive multi-domain collaborative denoising and enhancement strategy, which makes full use of domain knowledge and prior information to improve the generalization ability and processing effect of the method.

[0036] This application adaptively selects denoising and enhancement strategies based on noise detection and classification results, performs targeted processing on sensor images, and outputs high-quality image results.

[0037] The method provided in this application can effectively adapt to complex noise environments, achieve high-fidelity denoising and detail restoration in complex scenes, and improve image quality and texture preservation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flowchart of the denoising and enhancement processing method for sensor image data provided by this application;

[0039] Figure 2This is a high-fidelity denoising and enhancement strategy diagram for the denoising and enhancement processing method for sensor image data in complex noise scenarios provided by this application;

[0040] Figure 3 This is a diagram of the adaptive multi-domain collaborative denoising and enhancement strategy for the denoising and enhancement processing method for sensor image data provided in this application;

[0041] Figure 4 This is a structural diagram of the sensor image data denoising and enhancement processing system provided in this application. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the drawings in the specification.

[0043] In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without violating the connotation of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present application. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0045] Example 1

[0046] Reference Figures 1 to 3 The first embodiment of the present application provides a denoising and enhancement processing method for sensor image data.

[0047] Step 1: Obtain image data collected by the sensor and preprocess the image data.

[0048] Assume that the original image collected by the sensor is ,in Represents the row and column coordinates of the image respectively, and the pixel value range is ; Get the image data collected by the sensor and convert it into grayscale image data :

[0049]

[0050] in, Represents the pixel values of the red, green, and blue color channels of the image data respectively.

[0051] Grayscale image data Normalize the pixel value range to : ;in, Represents the normalized image data.

[0052] The Gaussian smoothing algorithm is used to normalize the image data Perform smoothing to reduce high-frequency noise in the image; the Gaussian smoothing operator is defined as:

[0053]

[0054] in, is the image data processed by the Gaussian smoothing operator, Represents the standard deviation of the Gaussian function, which controls the degree of smoothing. Represents an exponential function with a natural constant as its base.

[0055] Smoothed image data for: ;in, Represents the convolution operation.

[0056] Use Laplace operator to smooth the image data Perform sharpening to enhance image details; the Laplace operator is defined as:

[0057]

[0058] in, is the image data processed by the Laplace operator, Indicates the second-order partial derivative; It means to find the first-order partial derivative.

[0059] Sharpened image data for: ;in, Indicates the sharpening strength and controls the degree of sharpening.

[0060] After sharpening the image data Quantize the pixel value range and remap it to : ;in, Indicates floor operation, Represents the quantized image data.

[0061] In this step, the quantized image data As the preprocessed noisy image data, it is input into the subsequent steps for noise detection, classification, and denoising and enhancement processing; through the above preprocessing operations, the initial noise in the image can be effectively reduced, the image quality can be improved, and the foundation for subsequent denoising and enhancement processing can be laid.

[0062] Step 2: Build a deep learning model to detect the noise type of the preprocessed image data and determine the image denoising and enhancement strategy based on the noise type.

[0063] Construct a convolutional neural network (CNN) model for noise detection and classification. The model input is the preprocessed noise image data. , the output is the probability distribution of the noise type.

[0064] The structure of the CNN model includes: input layer, which is used to input preprocessed noisy image data ; In the first convolutional layer, the convolution kernel size is set to 3×3, the number of convolution kernels is 16, and the activation function is ReLU; in the first pooling layer, the maximum pooling is used, the pooling kernel size is 2×2, and the step size is 2; in the second convolutional layer, the convolution kernel size is set to 3×3, the number of convolution kernels is 32, and the activation function is ReLU; in the second pooling layer, the maximum pooling is used, the pooling kernel size is 2×2, and the step size is 2; in the first fully connected layer, the number of neurons is set to 128, and the activation function is ReLU; in the second fully connected layer, the number of neurons is set to , the activation function is Softmax.

[0065] Using an image dataset with noise type labels, the CNN model is trained using the cross entropy loss function and the Adam optimization algorithm; the cross entropy loss function Defined as:

[0066] ;

[0067] in, Indicates the The samples belong to True label of class noise (0 or 1), is the sample size, is the number of noise types, Indicates that the CNN model predicts The samples belong to The probability of noise.

[0068] Use the trained CNN model to classify noise and preprocess the noise image data Input the trained CNN model to obtain the probability distribution of noise types, and select the category with the highest probability as the final noise type.

[0069] According to the detected noise type, the corresponding denoising and enhancement strategy is selected to achieve the optimal image quality improvement effect.

[0070] Step 3: Construct a high-fidelity denoising and enhancement strategy for complex noise scenes, including noise domain decomposition, dynamic area weighted denoising, and multi-scale enhancement and texture compensation.

[0071] Step 301: Noise domain decomposition.

[0072] For the preprocessed noisy image Perform wavelet transform to decompose the image into low-frequency subbands and high frequency sub-band 、 、 :

[0073] ;

[0074] in, and represent low-pass filter and high-pass filter respectively, represents the number of decomposition levels, and represents the offset of the filter on the image, Indicates Axis offset, Indicates The offset of the axis.

[0075] For high frequency sub-band 、 、 Perform adaptive threshold denoising, threshold According to the noise standard deviation Adaptive Adjustment:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] in, 、 、 represents the high-frequency subband after denoising, Indicates the width of the image, Indicates the length of the image.

[0081] For low frequency sub-band and the denoised high frequency subband 、 、 Perform inverse wavelet transform and reconstruct the denoised image :

[0082] .

[0083] Step 302: Dynamic region weighted denoising.

[0084] Denoised image Divide into sizes Overlapping blocks ,in and Represent the row and column coordinates of the block respectively.

[0085] For each overlapping block , search for similar overlapping blocks in its neighborhood and build a set of similar blocks Similar overlapping blocks are determined using Euclidean distance. A Euclidean distance threshold is set between blocks. If the distance between blocks is less than the Euclidean distance threshold, the blocks are considered similar overlapping blocks. The calculation formula for the Euclidean distance d between overlapping blocks is:

[0086] ;

[0087] in, Represents candidate similar blocks in the search neighborhood.

[0088] For each similar block set , perform collaborative denoising to obtain the denoised overlapping blocks : ;in, Indicates the size of the set of similar blocks.

[0089] Perform weighted averaging on the denoised overlapping blocks to obtain the final denoised image :

[0090] ;

[0091] in, Represents overlapping blocks after denoising In coordinates The weight at can be calculated using Gaussian weights:

[0092] ;

[0093] in, Represents overlapping blocks after denoising The center coordinates of represents the standard deviation of the Gaussian kernel.

[0094] Step 303: Multi-scale enhancement and texture compensation.

[0095] The final denoised image Perform multi-scale decomposition to obtain sub-band coefficients at different scales :

[0096] ;

[0097] in, represents the number of decomposition levels, Represents a low-pass filter, which is obtained by adjusting the sub-band coefficients of the previous level Perform low-pass filtering and downsampling to obtain the current scale layer The subband coefficients .

[0098] For each scale subband coefficient Enhance and obtain the enhanced subband coefficients : ;in, Indicates the The enhancement factor of the layer can be adaptively adjusted according to the noise type and intensity.

[0099] On the basis of multi-scale enhancement, texture compensation is performed on the denoised high-frequency sub-band to restore the texture details of the image. Texture compensation can adopt a sparse representation-based method, such as K-SVD dictionary learning: ;in, Indicates the The first high frequency subband The blocks here divide the high-frequency sub-band into several small blocks, and each block performs independent dictionary learning and sparse coding to better capture local texture features. Represents a dictionary, represents the sparse coding coefficient, represents the sparsity regularization parameter.

[0100] The enhanced sub-band coefficients and the high-frequency sub-bands after texture compensation are reconstructed at multiple scales to obtain the final enhanced image. ;in, Represents the texture compensation Layer high frequency subband.

[0101] This step can achieve high-fidelity image denoising and enhancement in complex noise scenes through noise domain decomposition, dynamic area weighted denoising, multi-scale enhancement and texture compensation, effectively improving the visual quality and signal-to-noise ratio of the image.

[0102] Step 4: Construct an adaptive multi-domain collaborative denoising and enhancement strategy, including noise domain adaptive projection based on the variational autoencoder model, multi-scale detail enhancement under the pyramid structure, and cross-domain texture migration and fusion.

[0103] Step 401: Noise domain adaptive projection based on variational autoencoder.

[0104] Construct a variational autoencoder model to map the noise image to the latent space and realize adaptive projection in the noise domain; the VAE (variational autoencoder) model consists of an encoder and decoder Composition, of which and Represent the parameters of the encoder and decoder respectively.

[0105] encoder The preprocessed noisy image Mapping to latent variables The posterior distribution of , assuming the posterior distribution is Gaussian distribution: ;in, and Represent the mean and variance of the posterior distribution, respectively, output by the encoder network; the decoder From the hidden variable Generate a reconstructed image.

[0106] The objective function of the VAE model Including reconstruction loss and KL divergence regularization term:

[0107] ;

[0108] in, represents the likelihood function of the decoder, Represents latent variables The prior distribution of is usually assumed to be a standard normal distribution ; Train the VAE model according to the objective function of the VAE model.

[0109] Using the trained VAE model, the preprocessed noise image Mapping to the latent space to obtain an adaptive noise representation : ; Represents the encoding function The preprocessed noisy image Mapped to the latent space, according to the noise representation , generate the corresponding denoised image : .

[0110] Step 402: Multi-scale detail enhancement under a pyramid structure.

[0111] Denoised image Perform Gaussian pyramid decomposition to obtain image pyramids at different scales ,in Indicates the pyramid layer index, Represents the total number of pyramid layers; the Gaussian pyramid decomposition process is expressed as: ;in, Indicates the layers of the image pyramid, Indicates the layers of the image pyramid, The standard deviation is Gaussian filter, represents the convolution operation, Represents a downsampling operation.

[0112] Construct a Laplacian pyramid based on image pyramids at different scales , used to extract multi-scale detail information of images: ;in, Represents an upsampling operation.

[0113] Adaptively enhance each layer of the Laplace pyramid to obtain the enhanced Laplace pyramid : ;in, Indicates the The Laplacian pyramid of layers, Indicates the The enhancement coefficient of the layer can be adaptively adjusted according to the noise level.

[0114] Reconstruct the enhanced Laplacian pyramid to obtain an image with enhanced details .

[0115] Step 403: Cross-domain texture migration and fusion.

[0116] From a reference image without noise Extract reference texture features to guide texture migration of noisy images; use Gabor filter or wavelet transform method to extract multi-scale and multi-directional texture features: ;in, represents the extracted reference texture features, Represents the feature extraction function.

[0117] Using reference texture features , for the image after detail enhancement Perform texture migration to generate texture-enhanced images ; Use a texture synthesis method based on deep convolutional neural network to generate texture-enhanced images: ,in, is the texture synthesis function.

[0118] Texture-enhanced image With denoised image Perform multi-domain fusion to obtain the final enhanced image : ;in, Represents the weight of texture enhancement, which can be adaptively adjusted according to the noise level.

[0119] This step can achieve adaptive multi-domain collaborative denoising and enhancement through noise domain adaptive projection based on variational autoencoder, multi-scale detail enhancement under the pyramid structure, and cross-domain texture migration and fusion, effectively improving the visual quality and detail fidelity of the image.

[0120] Step 5: According to the noise detection and classification results, select the corresponding denoising and enhancement strategy, perform denoising and enhancement on the sensor image, and output the processed sensor image.

[0121] According to the noise detection and classification results of the convolutional neural network CNN model on the image, the mapping relationship between noise type and denoising and enhancement strategy is established; let the noise type be ,in , Represents different noise types; denoising and enhancement strategies include high-fidelity denoising and enhancement strategies in complex noise scenes and adaptive multi-domain collaborative denoising and enhancement strategy .

[0122] Define the mapping function between noise type and denoising and enhancement strategies :

[0123] ;

[0124] in, Indicates the denoising and enhancement strategy applicable to the j-th noise type, Indicates that it applies to the strategy A subset of noise types, Indicates that it applies to the strategy The noise type subset is determined based on prior knowledge or data analysis.

[0125] For a given preprocessed noisy image , through the trained convolutional neural network CNN model, the probability distribution of noise type is obtained ,in Indicates that the image is of noise type probability.

[0126] According to the probability distribution , select the noise type with the highest probability ; Using the mapping function , determine the corresponding denoising and enhancement strategies : .

[0127] According to the selected denoising and enhancement strategy , for the preprocessed noisy image Perform denoising and enhancement processing.

[0128] if , a high-fidelity denoising and enhancement strategy is adopted in complex noise scenes, including the following steps:

[0129] Noise domain decomposition: The noisy image is subjected to wavelet transform to decompose it into low-frequency sub-bands and high-frequency sub-bands, and adaptive threshold denoising is performed on the high-frequency sub-bands;

[0130] Dynamic area weighted denoising: Divide the denoised image into overlapping blocks, perform similar block search and weighted averaging on each overlapping block to obtain the denoised image;

[0131] Multi-scale enhancement and texture compensation: The denoised image is decomposed into multiple scales, the sub-band coefficients of each scale layer are enhanced, and texture compensation is performed on the high-frequency sub-bands. Finally, the enhanced image is reconstructed.

[0132] if , an adaptive multi-domain collaborative denoising and enhancement strategy is adopted, including the following steps:

[0133] Noise domain adaptive projection based on variational autoencoder: Using the trained variational autoencoder (VAE) model, the noisy image is mapped to the latent space to obtain an adaptive noise representation, and then the corresponding denoised image is generated based on the noise representation;

[0134] Multi-scale detail enhancement under the pyramid structure: The denoised image is decomposed into a Gaussian pyramid, a Laplacian pyramid is constructed, each layer is adaptively enhanced, and finally the image with detail enhancement is reconstructed;

[0135] Cross-domain texture transfer and fusion: Extract reference texture features from the reference image, perform texture transfer on the detail-enhanced image, generate a texture-enhanced image, and then perform multi-domain fusion with the original denoised image to obtain the final enhanced image.

[0136] Output the corresponding processing results according to the selected denoising and enhancement strategies :

[0137] ;

[0138] The processed image As the output result of sensor image denoising and enhancement, it is used for subsequent applications or presentations.

[0139] This step adaptively selects appropriate denoising and enhancement strategies based on the noise detection and classification results, and performs targeted denoising and enhancement processing on the sensor image.

[0140] The technical solution of this application can flexibly select corresponding processing strategies according to different noise types to improve the effect and adaptability of denoising and enhancement; at the same time, through the organic combination of noise domain decomposition, dynamic area weighting, multi-scale enhancement and cross-domain texture migration technology, high-fidelity, high-quality image restoration and enhancement can be achieved in complex noise scenes, providing reliable image data support for related image usage scenarios.

[0141] Example 2

[0142] Reference Figure 4 The second embodiment of the present application provides a denoising and enhancement processing system for sensor image data.

[0143] The system includes: a data acquisition module, a noise detection module, a first denoising and enhancement module, a second denoising and enhancement module and an image output module.

[0144] The data acquisition module is used to acquire image data collected by the sensor and pre-process the image data.

[0145] The noise detection module is used to build a deep learning model, detect the noise type of the preprocessed image data, and determine the image denoising and enhancement strategy according to the noise type.

[0146] The first denoising and enhancement module is used to construct a high-fidelity denoising and enhancement strategy in complex noise scenes, including noise domain decomposition, dynamic area weighted denoising, and multi-scale enhancement and texture compensation.

[0147] The second denoising and enhancement module is used to construct an adaptive multi-domain collaborative denoising and enhancement strategy, including noise domain adaptive projection based on a variational autoencoder model, multi-scale detail enhancement under a pyramid structure, and cross-domain texture migration and fusion.

[0148] The image output module selects corresponding denoising and enhancement strategies according to noise detection and classification, performs denoising and enhancement processing on the sensor image, and outputs the processed sensor image.

[0149] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0150] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose of this application and the scope of protection of the claims, which are all within the protection of this application.

Claims

1. A method for denoising and enhancing sensor image data, characterized in that: include: Acquire image data collected by the sensor and preprocess the image data; Build a deep learning model to detect the noise type of preprocessed image data and determine the image denoising and enhancement strategy based on the noise type; Build high-fidelity denoising and enhancement strategies for complex noise scenarios, including noise domain decomposition, dynamic region weighted denoising, and multi-scale enhancement and texture compensation; The dynamic area weighted denoising method includes: dividing the denoised image into overlapping blocks; for each overlapping block, searching for similar overlapping blocks in the neighborhood to construct a similar block set, wherein similar overlapping blocks are determined using Euclidean distance; performing collaborative denoising on each similar block set to obtain denoised overlapping blocks; and performing weighted averaging on the denoised overlapping blocks to obtain a final denoised image. The multi-scale enhancement and texture compensation includes: performing multi-scale decomposition on the final denoised image to obtain sub-band coefficients at different scales; enhancing the sub-band coefficients at each scale to obtain enhanced sub-band coefficients; performing texture compensation on the denoised high-frequency sub-band based on the multi-scale enhancement to restore the texture details of the image; and performing multi-scale reconstruction on the enhanced sub-band coefficients and the texture-compensated high-frequency sub-band to obtain an enhanced image. Constructing adaptive multi-domain collaborative denoising and enhancement strategies, including noise domain adaptive projection based on variational autoencoder models, multi-scale detail enhancement under pyramid structures, and cross-domain texture transfer and fusion; The variational autoencoder model includes: a method for mapping a noisy image to a latent space to achieve adaptive projection in the noise domain; the variational autoencoder model is composed of an encoder and a decoder; the encoder maps the noisy image to a posterior distribution of a latent variable; the decoder generates a reconstructed image from the latent variable; the objective function of the variational autoencoder model includes a reconstruction loss and a KL divergence regularization term, and the variational autoencoder model is trained using the objective function; the noisy image is mapped to the latent space using the trained variational autoencoder model to obtain an adaptive noise representation; and the corresponding denoised image is generated based on the noise representation; The cross-domain texture migration and fusion includes: extracting reference texture features from a reference image to guide texture migration of the noisy image; using the reference texture features to perform texture migration on the detail-enhanced image to generate a texture-enhanced image; and performing multi-domain fusion of the texture-enhanced image with the original denoised image to obtain a final enhanced image. According to noise detection and classification, the corresponding denoising and enhancement strategies are selected, the sensor image is denoised and enhanced, and the processed sensor image is output.

2. The method for denoising and enhancing sensor image data according to claim 1, wherein: Acquire image data collected by the sensor, perform grayscale conversion, normalization, image data smoothing, image data sharpening, and image data quantization on the acquired image data; after pre-processing, the image data is used for noise type detection; Construct a convolutional neural network (CNN) model for noise detection and classification. The model input is preprocessed noise image data, and the output is the probability distribution of noise types. The convolutional neural network (CNN) model includes an input layer, a convolution layer, a pooling layer, and a fully connected layer. The CNN model is trained using an image dataset with noise type labels, a cross entropy loss function, and an Adam optimization algorithm. The trained CNN model is used to classify noise. The preprocessed noise image data is input into the trained CNN model to obtain the probability distribution of noise types, and the category with the highest probability is selected as the final noise type.

3. The method for denoising and enhancing sensor image data according to claim 2, wherein: The noise domain decomposition includes: performing wavelet transform on the noisy image to decompose the image into low-frequency sub-bands and high-frequency sub-bands, performing adaptive threshold denoising on the high-frequency sub-bands, and adaptively adjusting the threshold according to the noise standard deviation; performing inverse wavelet transform on the low-frequency sub-bands and the denoised high-frequency sub-bands to reconstruct the denoised image.

4. The method for denoising and enhancing sensor image data according to claim 3, wherein: The multi-scale detail enhancement under the pyramid structure includes: performing Gaussian pyramid decomposition on the denoised image to obtain image pyramids at different scales; constructing a Laplacian pyramid based on the image pyramids at different scales to extract multi-scale detail information of the image; adaptively enhancing each layer of the Laplacian pyramid to obtain an enhanced Laplacian pyramid; and reconstructing the enhanced Laplacian pyramid to obtain a detail-enhanced image.

5. The method for denoising and enhancing sensor image data according to claim 4, wherein: For a given noise image, the probability distribution of noise types is obtained through the trained convolutional neural network (CNN) model. According to the probability distribution, the noise type with the highest probability is selected. Based on the noise detection and classification results, a mapping relationship between noise type and denoising and enhancement strategies is established; According to the mapping relationship, the corresponding denoising and enhancement strategy is selected, the noisy image is denoised and enhanced, and the corresponding processing results are output.

6. A sensor image data denoising and enhancement processing system, which is used to implement the sensor image data denoising and enhancement processing method according to any one of claims 1 to 5, characterized in that: include: Data acquisition module, noise detection module, first denoising and enhancement module, second denoising and enhancement module and image output module; The data acquisition module is used to acquire image data collected by the sensor and pre-process the image data; The noise detection module is used to build a deep learning model, detect the noise type of the preprocessed image data, and determine the image denoising and enhancement strategy according to the noise type; The first denoising and enhancement module is used to construct a high-fidelity denoising and enhancement strategy in complex noise scenes, including noise domain decomposition, dynamic area weighted denoising, and multi-scale enhancement and texture compensation; The second denoising and enhancement module is used to construct an adaptive multi-domain collaborative denoising and enhancement strategy, including noise domain adaptive projection based on a variational autoencoder model, multi-scale detail enhancement under a pyramid structure, and cross-domain texture migration and fusion; The image output module selects corresponding denoising and enhancement strategies according to noise detection and classification, performs denoising and enhancement processing on the sensor image, and outputs the processed sensor image.

Images