A low-illumination image enhancement method under a freight security monitoring scene

Abstract

The application discloses a low-illumination image enhancement method suitable for freight security monitoring scenes. The method first collects low-illumination original image data through a high-definition starlight-level camera, and carries out non-local mean filtering and linear normalization preprocessing; further, an improved discrete cosine transform filter bank is used for frequency domain transformation of the image, a filter bank with direction selectivity is designed based on image frequency characteristics, texture complexity and edge characteristics, so that multi-frequency domain and multi-direction features are extracted, and separation of noise and effective details is realized; subsequently, multi-source image features are adaptively fused through a fusion rule based on contrast weighting; finally, the fused features are input into a specially trained encoder-decoder structure convolutional neural network, and optimization training is carried out by using a multi-component loss function, so that deep enhancement and reconstruction of the image are completed.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to a low-light image enhancement method for freight security monitoring scenarios, which aims to improve the clarity and visibility of freight monitoring images under low light conditions. Background Technology

[0002] In the field of freight security monitoring, ensuring the clarity and visibility of surveillance images is a crucial aspect of safeguarding logistics transportation safety. However, in practical applications, due to the uncertainty of lighting conditions, especially at night or in low-light environments, images captured by surveillance cameras often suffer from low illumination. Low-illumination images not only result in insufficient overall brightness but also cause a series of problems such as decreased contrast, loss of detail, and increased noise, severely impacting monitoring effectiveness and subsequent image analysis and processing. Traditional low-illumination image enhancement methods mainly rely on simple brightness adjustments and contrast stretching, such as histogram equalization and gamma correction. While these methods can improve image brightness to some extent, they often fail to effectively recover lost detail information and may even amplify noise in the image, resulting in unsatisfactory enhanced image quality. Furthermore, traditional methods typically lack adaptability to image content and are difficult to optimize for different scenes and lighting conditions.

[0003] With the rapid development of computer vision and deep learning technologies, data-driven image enhancement methods have gradually become a research hotspot. These methods learn the mapping relationship for image enhancement from a large amount of labeled data, enabling them to automatically adapt to low-light conditions in different scenarios and achieve more accurate and effective image enhancement. However, existing deep learning methods still face many challenges when processing low-light images. For example, many methods are highly dependent on training data, requiring a large number of labeled low-light-normal-light image pairs for training, which is often difficult to meet in practical applications. Simultaneously, some methods are prone to color distortion and detail blurring when processing images under complex lighting conditions, making it difficult to meet the high image quality requirements of freight safety monitoring. Therefore, there is an urgent need for an image enhancement method and system that can effectively improve image brightness, preserve and restore image details, and has good adaptability to different lighting conditions, addressing the problem of low-light image enhancement in freight safety monitoring scenarios. This invention aims to overcome the shortcomings of existing technologies by providing an efficient and reliable low-light image enhancement solution for freight safety monitoring through multi-scale feature extraction, illumination estimation and compensation, and detail enhancement steps. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems of noise amplification, detail loss, color distortion, and low processing efficiency in existing low-light image enhancement methods, and to provide a low-light image enhancement method for freight safety monitoring scenarios. This method combines an improved Discrete Cosine Transform (DCT) filter bank with Convolutional Neural Networks (CNNs) to construct a complete technical solution from data preprocessing, noise separation, feature fusion to final image reconstruction. This solution can effectively suppress noise, preserve image details and true colors, and significantly improve the overall visual effect and usability of the image, thereby meeting the high standards of image clarity and reliability required for freight monitoring.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Figure 1 As shown, the core process of this method specifically includes the following steps:

[0006] Step S1, Image Data Acquisition: High-definition cameras deployed in monitoring areas such as freight warehouses, logistics transfer stations, container yards, and freight channels are used to acquire raw image data in low-light environments. These cameras preferably use low-light CMOS sensors or starlight-level cameras, specifically designed for operation in environments such as nighttime, dusk, and poorly lit indoor spaces. The raw images acquired typically suffer from low overall brightness, poor contrast, significant noise, and blurred details.

[0007] Step S2, Data Preprocessing: The raw image data acquired in Step S1 is preprocessed to lay the foundation for subsequent depth processing. Preprocessing operations mainly include denoising and normalization. Denoising uses spatial or frequency domain filtering methods to initially suppress noise in the raw image, reducing the interference of random noise on subsequent transformations and analyses. Normalization performs a linear mapping on the denoised image data, uniformly converting pixel values ​​to a specific numerical range to eliminate the influence of dimensions, accelerate the convergence of subsequent processing, and improve numerical stability. Specifically:

[0008] Let the original image data be I = (i1, i2, ..., i... m The normalized image data is I′=(i′1,i′2,…,i′) , where m and n are the number of rows and columns of the image, respectively. m×n The calculation method is as follows:

[0009]

[0010] Where min(I) and max(I) are the minimum and maximum values ​​of pixels in image I, respectively.

[0011] Step S3: Frequency Domain Transformation and Component Separation Based on Improved DCT Filter Bank: For the preprocessed image data I', a specially designed improved DCT filter bank is used for frequency domain analysis to separate noise components from useful image details. This step, through a strategy of separation before enhancement, effectively avoids the common problem of simultaneous noise and detail amplification in traditional enhancement methods.

[0012] First, suitable DCT basis functions should be selected or constructed based on the characteristics of low-light images. The selection of basis functions requires comprehensive consideration of several factors: the distribution of image energy in the frequency domain needs to be analyzed, and the selection should be optimized based on the proportion of low-frequency and high-frequency information; the demand for high-frequency component representation in texture-rich regions needs to be considered; the sensitivity of the basis functions to edge directions needs to be evaluated to better preserve edge information; the low contrast characteristic of low-light images needs to be analyzed, and basis functions that can enhance specific frequency components to improve contrast need to be selected; and computational efficiency needs to be considered while ensuring performance to meet the real-time requirements of the monitoring system.

[0013] Secondly, based on the selected basis functions and the frequency domain characteristics of the low-light image, an improved DCT filter bank including low-pass and high-pass filters is designed. The design process includes: determining the appropriate DCT transform order according to the image size, detail richness, and processing accuracy requirements; solving for the zeros of the polynomials corresponding to the selected DCT basis using mathematical methods; using the solved zeros, calculating the specific coefficients of the low-pass and high-pass filters using spectral decomposition or filter design algorithms; and finally, verifying through calculation whether the designed filter bank meets the requirements of orthogonality and completeness, ensuring that the original signal can be reconstructed without information loss.

[0014] Next, the preprocessed image I' is transformed using a pre-designed improved DCT filter bank, converting the image from the spatial domain to the frequency domain, resulting in a set of DCT coefficients. By analyzing these coefficients distributed across different frequency subbands and directions, multi-frequency features and directional features can be extracted. Multi-frequency feature analysis can separate low-frequency coefficients representing the overall illumination and contours of the image, mid-frequency coefficients representing the main textures and edges, and high-frequency coefficients representing fine details and noise; directional feature analysis utilizes the directional sensitivity of the two-dimensional DCT to analyze the distribution of coefficients in different directions to distinguish edges and textures with different orientations.

[0015] Finally, rules are formulated to selectively retain or suppress DCT coefficients of specific frequencies and directions based on the specific monitoring application requirements. By setting appropriate thresholds to attenuate small amplitude values ​​in high-frequency coefficients, while retaining or enhancing components corresponding to important edges in mid-frequency coefficients, effective separation of noise components from useful image details is achieved.

[0016] Step S4: Data fusion processing; Due to the possible use of multiple cameras or multi-scale, multi-directional image decomposition, multiple sets of transform coefficients will be obtained. The goal of step S4 is to effectively fuse these separated, complementary, or redundant image data to generate a set of fused transform coefficients C that more comprehensively and accurately reflects scene information. new This invention employs a contrast-weighted fusion rule: First, for the two sets of transform coefficients C of images A and B that need to be fused... A and C B Calculate their contrast values ​​respectively. Contrast can be defined as the variance of coefficient values ​​within a local region, the sum of gradient magnitudes, or other measures that reflect information richness and sharpness. Let their contrast values ​​be C... A and C B Then, a fusion weight is assigned to each group of coefficients based on contrast. The weight calculation formula is as follows:

[0017]

[0018] This weighting strategy ensures that coefficients with high contrast and richer information have a greater weight in the fusion process. The weighted sum of the two sets of coefficients is then used to obtain the new fused coefficient C. new :

[0019] C new =W A C A +W B C B

[0020] Finally, an inverse discrete cosine transform is performed on the fused coefficients to reconstruct the fused image. This image combines the advantages of multi-source information, resulting in better clarity and information integrity.

[0021] Step S5, CNN-based image enhancement and reconstruction: The fused transform coefficients C obtained in step S4 are... new Alternatively, the fused image, or its reconstructed version, can be used as input and fed into a specially designed CNN-based image enhancement model for end-to-end optimization and reconstruction. This model achieves image enhancement by learning a complex nonlinear mapping from low-quality fused features to high-quality, sharp images.

[0022] The model structure includes an input layer, convolutional layers, feature extraction layers, hidden layers, and an output layer. The input layer receives the fused transform coefficient matrix or reconstructed image patches. The convolutional layers consist of multiple convolutional kernels, performing multi-scale convolution operations on the input. By selecting different kernel sizes and setting an appropriate number of convolutional layers, a deep feature extraction network is constructed. The feature extraction layer is essentially a collection of feature maps output from the convolutional layers, further processed, including pooling layers to increase the receptive field and introducing non-linear activation functions to enhance the model's non-linear expressive power. The hidden layers typically consist of fully connected layers or deeper convolutional layers, responsible for performing non-linear transformations, combinations, and adjustments on the features output from the feature extraction layers. The output layer is usually a convolutional or deconvolutional layer, generating the final enhanced and sharpened image based on the high-level features learned by the hidden layers.

[0023] The input parameters used during model training mainly include the original low-light image dataset, the convolutional kernel function defined in the model structure, and the number of convolutional layers and the number of channels per layer. The model output parameters include the enhanced, sharpened image directly output during inference, and optional quantitative metrics for objectively evaluating the degree of image quality improvement. By training on large-scale paired datasets, this CNN model can learn complex enhancement priors, effectively recover image details, adjust global and local contrast, and maintain color fidelity.

[0024] To meet the needs of networked monitoring and remote processing, the method also includes two stages: compression encoding and transmission, and decoding recovery and enhancement. Before transmitting the raw or intermediate processed data over the network, the sparsity of the coefficients after the DCT transform in step S3 is utilized to compress the data using entropy coding or standard image compression techniques, reducing transmission bandwidth consumption. At the receiving end, the compressed bitstream is decoded, and the image data is recovered through inverse DCT transform. The recovered image is then fed into the CNN enhancement model in step S5 for secondary enhancement processing to further improve the quality of the final output image.

[0025] Compared with existing technologies, this invention has the following significant advantages: the method is specifically designed for low-light scenarios in freight monitoring, taking into account the unique defects of this type of image; by improving the DCT filter bank to separate noise from useful components in advance, the noise amplification problem that may occur when CNN directly processes the image is avoided; a multi-source, multi-feature fusion mechanism is introduced to comprehensively utilize the complementary advantages of different information sources; by utilizing the powerful nonlinear fitting capability of CNN, complex brightness adjustment, detail restoration, and color correction can be achieved; a complete system solution from acquisition, processing to display is formed, and network transmission optimization is considered, which facilitates practical deployment and application.

[0026] This invention effectively enhances the perception capability of the freight safety monitoring system under adverse lighting conditions, which is of great significance for ensuring cargo safety and improving the efficiency of warehousing and logistics management. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the method of the present invention; Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are merely illustrative and not intended to limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art based on the technical solutions of the present invention should be included within the protection scope of the present invention.

[0029] Example 1: Low-light image enhancement method in freight safety monitoring scenarios

[0030] This embodiment provides a complete implementation process for a low-light image enhancement method, which is particularly suitable for low-light monitoring scenarios such as large logistics parks, port freight areas, and nighttime warehouses.

[0031] System hardware configuration

[0032] Before implementing this method, a corresponding hardware platform needs to be set up. Within the freight monitoring area, high-definition, low-light network cameras are deployed at intervals of 50-100 meters. These cameras are starlight-level cameras with 1 / 1.8-inch CMOS sensors, a minimum illumination of 0.001 lux, support H.265 encoding, and a resolution of 1920×1080 or 2560×1440. All cameras are connected to an image processing server deployed in the monitoring center via a network switch. This server is equipped with an Intel Xeon processor, 128GB of memory, and an NVIDIA RTX 6000 Ada GPU to run image processing algorithms and neural network models. The display terminal is a 46-inch 4K monitoring screen array.

[0033] Method Implementation Steps

[0034] Step S1: Image Data Acquisition

[0035] After the system starts, cameras at each monitoring point continuously capture video streams of the monitored area at a preset frame rate (usually 15-25fps). At night or during periods of low light, the cameras automatically switch to low-light mode, increasing sensor gain and extending exposure time to obtain the clearest possible original image. Simultaneously, the system records metadata such as the timestamp, camera number, and location information for each frame. For example, in the eastern section of a logistics park, three cameras monitor the loading and unloading area from different angles. At 3 AM, with ambient light levels of only 1-2 lux, the average brightness value of the original images captured by the cameras is only 15-30 (the normal range is 0-255), and there is significant random noise and excessive darkness in some areas.

[0036] Step S2: Data Preprocessing

[0037] The acquired raw images are first sent to the preprocessing module. Preprocessing includes two main operations:

[0038] Denoising: A Non-Local Means (NLM) filtering algorithm is used for initial image denoising. For each pixel i in the input image I, the NLM filtered value is obtained by calculating a weighted average of all pixels in the image, where the weights depend on the similarity between pixels i and j. The specific calculation formula is as follows:

[0039]

[0040] Among them, the weights w(i,j) satisfy 0≤w(i,j)≤1 and ∑ j w(i,j) = 1, and the weights are determined by the Gaussian-weighted Euclidean distance between image patches centered at pixels i and j:

[0041]

[0042] Here N k Represents a fixed-size neighborhood centered at pixel k, ||·|| 2,a Let represent the Gaussian-weighted Euclidean distance, h be the filtering parameter controlling the attenuation level, and Z(i) be the normalization constant. In practice, to balance denoising effect and computational efficiency, the search window is set to 21×21 pixels, the similarity block size is 7×7 pixels, and the filtering parameter h = 10. After NLM filtering, the Gaussian noise in the image is significantly suppressed, and the peak signal-to-noise ratio is improved by an average of about 3-5 dB.

[0043] Normalization: The denoised image is then normalized. Let the denoised image be I. denoised First, calculate the minimum pixel value I of the entire image. min and maximum pixel value I max Then for each pixel value Idenoised Perform a linear transformation on (x, y):

[0044]

[0045] After this transformation, all pixel values ​​are mapped to the [0, 1] interval. For color images, the above normalization operation is performed independently on the R, G, and B channels respectively. After denoising and normalization, the overall contrast of the image is improved, and details in the dark areas are revealed, providing a higher quality input for subsequent processing.

[0046] Step S3: Frequency Domain Transformation and Component Separation Based on Improved DCT Filter Bank

[0047] This step is one of the core innovations of the method, and its specific implementation is as follows:

[0048] (1) Basis function selection

[0049] Considering the characteristics of freight surveillance images, a suitable DCT basis function is selected. Given that freight surveillance images typically contain numerous straight edges and regular textures such as cargo outlines and vehicle boundaries, two-dimensional DCT is chosen as the fundamental transformation, and its basis function is:

[0050]

[0051] in, α n Similar to the definition, M and N represent the image patch size. In practical implementation, analysis of the spectral characteristics of a large number of freight monitoring image samples revealed that energy is mainly concentrated in the mid-to-low frequency region, while high-frequency components are relatively sparse but contain important edge information. Therefore, based on standard two-dimensional DCT, a frequency weighting adjustment mechanism was introduced to give higher weight to mid-frequency components in specific directions.

[0052] (2) Improved DCT filter bank design

[0053] a. Determine the transform order: Divide the image into 8×8 non-overlapping blocks for DCT transform. This size provides a good balance between computational complexity and frequency resolution.

[0054] b. Filter Coefficient Design: Based on the DCT basis functions, design a filter bank with four directional subbands: horizontal, vertical, main diagonal, and sub-diagonal. Each directional filter is obtained by directional modulation of the standard DCT basis functions. For example, the frequency response of the horizontal filter emphasizes the frequency components in the horizontal direction and suppresses the frequency components in the vertical direction.

[0055] c. Calculate filter coefficients: For each 8×8 image patch, calculate its DCT coefficients under different directional filters. Let the original image patch be B, and the transformation matrix of the k-th directional filter be T. k The corresponding DCT coefficient matrix is:

[0056]

[0057] Among them, T k It is an 8×8 transformation matrix constructed based on the DCT basis functions after direction modulation.

[0058] d. Verify orthogonality and completeness: Verify whether each transformation matrix satisfies orthogonality, i.e. And whether the synthesis of all directional filters can completely reconstruct the original image patch. Through numerical verification, the reconstruction error of the filter bank designed in this embodiment is less than 0.1%, which meets the requirements of practical applications.

[0059] (3) Frequency domain transformation and feature extraction

[0060] For each normalized image patch, the four directional filter banks designed above are applied to obtain four sets of DCT coefficient matrices C. k k = 1, 2, 3, 4. Each coefficient matrix contains 64 coefficients, corresponding to different spatial frequencies. First, each coefficient matrix is ​​divided into three frequency bands from low to high frequency: low frequency band, containing DC components and the lowest 6 AC components, mainly reflecting the average brightness and basic outline of image patches; mid frequency band, containing 20 mid-frequency components, corresponding to the main texture and edge information of the image; high frequency band, containing the remaining 38 high-frequency components, corresponding to the fine details and noise of the image. Next, by analyzing the magnitude of the coefficients of the same frequency under filters in different directions, the dominant direction of that frequency component is determined. For example, if the absolute value of the coefficient of a certain spatial frequency under a horizontal filter is significantly greater than that in other directions, then that frequency component is considered to mainly exhibit horizontal characteristics.

[0061] (4) Component separation

[0062] Based on the feature analysis results, a component separation strategy is implemented: for high-frequency band coefficients, a threshold T is set. h =0.05×max(|C k |), where the absolute value is less than T hThe coefficients are treated as noise and set to zero, while coefficients above a threshold are retained as potentially important details. For mid-frequency coefficients, a direction-selective enhancement strategy is employed: for each frequency position, the absolute values ​​of its coefficients in four directions are compared, and the coefficients in the direction with the maximum value are retained. Coefficients in other directions are appropriately attenuated by multiplying them by an attenuation factor of 0.3-0.5. This strategy enhances the edge features in the dominant direction. Low-frequency coefficients remain largely unchanged to maintain the basic structure and brightness distribution of the image. Through these operations, effective separation of noise components from useful image details is achieved.

[0063] Step S4: Data Fusion Processing

[0064] In freight monitoring scenarios, multiple cameras often monitor the same area, or the same camera captures images at different times. This step fuses this multi-source information:

[0065] For two images of the same scene from different cameras or at different time periods, the contrast of their DCT coefficient matrices after processing in step S3 is first calculated. Contrast is defined as the ratio of the standard deviation to the mean of the coefficient matrix.

[0066]

[0067] Where σ(C) is the standard deviation of the coefficient matrix C, μ(C) is the mean, and ε = 10. -6 To prevent division by zero by small constants. Let the contrast of images A and B be C, respectively. A and C B The fusion weights are:

[0068]

[0069] For the DCT coefficients at each frequency position, they are fused according to weights:

[0070] C new (i, j) = W A C A (i, j) + W B C B (i, j)

[0071] Where (i, j) represents the position index in the coefficient matrix.

[0072] The fused DCT coefficient matrix is ​​subjected to inverse DCT transformation to reconstruct the fused image. For each 8×8 block, the inverse transformation formula is:

[0073]

[0074] The reconstructed results of all image patches are stitched together to obtain a complete fused image.

[0075] Step S5: CNN-based image enhancement and reconstruction

[0076] This step uses deep learning techniques to perform final enhancement on the fused images. Network architecture design: A CNN specifically designed for low-light image enhancement is constructed, named the "Freight Low-Light Enhancement Network". The network adopts an encoder-decoder structure and includes:

[0077] Encoder section: Input layer, receiving 256×256×3 RGB image blocks; Convolutional block 1, including 2 convolutional layers, each with 64 3×3 filters, ReLU activation, and batch normalization; Max pooling layer, 2×2 window, stride 2; Convolutional block 2, including 2 convolutional layers, each with 128 3×3 filters, ReLU activation, and batch normalization; Max pooling layer, 2×2 window, stride 2; Convolutional block 3, including 3 convolutional layers, each with 256 3×3 filters, ReLU activation, and batch normalization.

[0078] Bridge layer: Convolutional block 4, which includes 3 convolutional layers, each with 512 3×3 filters, ReLU activation, and batch normalization.

[0079] Decoder section: Upsampling layer 1, 2×2 transposed convolution, 256 filters; skip connection 1, concatenated with the output of convolutional block 3; convolutional block 5, including 2 convolutional layers, each with 256 3×3 filters, ReLU activation, batch normalization; upsampling layer 2, 2×2 transposed convolution, 128 filters; skip connection 2, concatenated with the output of convolutional block 2; convolutional block 6, including 2 convolutional layers, each with 128 3×3 filters, ReLU activation, batch normalization; upsampling layer 3, 2×2 transposed convolution, 64 filters; skip connection 3, concatenated with the output of convolutional block 1; convolutional block 7, including 2 convolutional layers, each with 64 3×3 filters, ReLU activation, batch normalization.

[0080] Output layer: the final convolutional layer, containing three 1×1 filters with sigmoid activation, outputting an enhanced RGB image.

Claims

1. A low-light image enhancement method for freight security monitoring scenarios, characterized in that, Includes the following steps: Step S1, Image Data Acquisition: Acquire raw image data of the freight monitoring area in low-light environment using a high-definition camera; Step S2, Data Preprocessing: The original image data is preprocessed, including denoising and normalization operations, to obtain preprocessed image data; Step S3: Frequency domain transformation and component separation based on improved DCT filter bank: For the preprocessed image data, frequency domain transformation is performed using an improved Discrete Cosine Transform (DCT) filter bank to extract features of multiple frequencies and directions, and to separate noise components and blurred components in the image. Step S4, Data Fusion Processing: Fusion processing is performed on image data separated from different sources or different feature layers to obtain a set of fused transform coefficients; Step S5: Image enhancement and reconstruction based on CNN: Input the fused transform coefficients into an image enhancement model based on Convolutional Neural Networks (CNN) for optimization and reconstruction, and output a clear enhanced image.

2. The low-light image enhancement method according to claim 1, characterized in that, The normalization operation in step S2 is specifically as follows: Let the original image data be I = (i1, i2, ..., i...). m×n ), where m and n are the number of rows and columns of the image, respectively. The normalized image data is I′=(i′1,i ′ 2,…,i′ m×n The calculation method is as follows: Where min(I) and max(I) are the minimum and maximum values ​​of pixels in image I, respectively.

3. The low-light image enhancement method according to claim 1, characterized in that, Step S3 includes: (1) Select the basis functions of DCT based on the image frequency characteristics, texture complexity, edge characteristics, contrast characteristics and computational complexity; (2) Based on the selected basis functions and the frequency domain characteristics of the image, an improved DCT filter bank is designed, which includes a low-pass filter and a high-pass filter. (3) The preprocessed image data is transformed using the filter bank to extract multi-frequency features and directional features; (4) Select the transformation coefficients to be retained according to the application requirements to achieve separation of noise and fuzzy components.

4. The low-light image enhancement method according to claim 3, characterized in that, The improved DCT filter bank designed in step S3 (2) includes: (1) Select the order of DCT based on the characteristics of the image; (2) Solve for the zeros of the DCT polynomial; (3) Calculate the coefficients of the low-pass filter and the high-pass filter based on the zero point; (4) Verify the orthogonality and completeness of the filter bank.

5. The low-light image enhancement method according to claim 1, characterized in that, The fusion process in step S4 employs the following method: First, based on the contrast-weighted fusion rule, the weights of the two transformation coefficients are calculated: Where C A and C B These are the contrast values ​​of the two transformation coefficients. Then, a new transformation coefficient C is obtained through weighted summation. new : C new =W A C A +W B C B Finally, the fused image is reconstructed through inverse transform.

6. The low-light image enhancement method according to claim 1, characterized in that, The CNN-based image enhancement model in step S5 includes: (1) An input layer, used to receive the fused transform coefficients or the image reconstructed therefrom; (2) Convolutional layer, used to perform multi-scale convolution operations on the input to generate feature maps; (3) Feature extraction layer, used to extract feature vectors from feature map; (4) Hidden layers are used for nonlinear transformation of features and advanced feature extraction; (5) Output layer, used to generate the enhanced image.

7. The low-light image enhancement method according to claim 6, characterized in that, The input parameters of the image enhancement model include the original image dataset, the convolution kernel function, and the number of convolution layers; the output parameters include the enhanced image and a quantitative indicator used to evaluate the degree of image quality improvement.

8. The low-light image enhancement method according to claim 1, characterized in that, Also includes: Before data transmission, the image data is compressed and encoded using the sparsity of the image. After receiving the data, the compressed data is decoded and inversely transformed to restore the image, and then enhanced.

Images