Real-time weak light image enhancement method and system based on lightweight multi-scale shift convolutional network

Abstract

The invention relates to a real-time weak light image enhancement method based on a lightweight multi-scale shift convolutional network, which belongs to the technical field of image processing, and comprises the following steps: S1, constructing a real-time weak light image enhancement network model based on the lightweight multi-scale shift convolutional network, comprising the following steps: establishing a light-weight weak light image enhancement network based on a depth Autoencoder architecture, determining core composition of a feature encoder, a bottleneck layer, a feature decoder and a loss function supervision module, and defining key parameters of DSConv and MSRB and overall configuration of the network; s2, training the real-time weak light image enhancement network model to obtain an optimal model parameter; and S3, building a real-time weak light image enhancement network model adaptive to the vehicle-mounted embedded platform by using the optimal model parameters, and enhancing a weak light image acquired by a vehicle-mounted camera in real time.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a real-time low-light image enhancement method and system based on a lightweight multi-scale shift convolutional network. Background Technology

[0002] With the popularization of assisted driving technology, lightweight low-light image enhancement algorithms for automotive embedded platforms have become a research hotspot. Research in this field mainly revolves around two core directions: lightweight algorithm design and optimization of basic module structure. In assisted driving scenarios, images captured by automotive cameras in low-light environments such as nighttime and tunnels are prone to color distortion and detail loss, severely impacting the execution of downstream visual tasks. However, automotive embedded platforms have limited computing power and resources, making it difficult for traditional low-light image enhancement schemes to balance processing efficiency and enhancement quality. Therefore, the development of lightweight algorithms is an urgent practical need.

[0003] In the field of low-light image enhancement for automotive embedded platforms, existing technologies have significant shortcomings in both lightweight algorithm design and basic module structure optimization. Traditional lightweight algorithm designs, whether supervised methods based on backbone networks like CNNs and Transformers, or unsupervised methods relying on deconvolutional networks and U-Net architectures, suffer from massive parameter counts due to model generality. Adding new branches or structures to adapt to automotive platforms further exacerbates deployment challenges. While basic lightweight designs can significantly reduce parameter counts, their ability to handle color distortion, high noise, and detail loss in complex scenarios is extremely limited. Subsequent lightweight algorithms incorporating more deep learning technologies often struggle to balance computational efficiency and inference performance, resulting in poor performance in real-time applications on automotive edge devices.

[0004] From the perspective of optimizing the basic module structure, fundamental paradigms such as grouped convolution and depthwise separable convolution only focus on reducing the number of parameters, neglecting the issues of accuracy degradation and information redundancy, and thus cannot meet the image quality requirements of automotive applications. While subsequent improvements have alleviated these problems to some extent, such as introducing channel shuffling and optimizing the grouping structure, the adaptability and optimization stability of convolution kernels remain insufficient in complex scenarios. Although the introduction of some dynamic or optimization mechanisms has improved the efficiency and robustness of convolution kernels in complex scenarios, the balance between module performance and computational cost remains prominent in practical applications, making it difficult to meet the resource-constrained characteristics of automotive embedded platforms and their stringent requirements for processing speed and image enhancement quality. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a real-time low-light image enhancement method and system based on a lightweight multi-scale shift convolutional network.

[0006] To achieve the above objectives, the present invention provides the following technical solution: On one hand, this invention provides a real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network, comprising the following steps: S1: Construct a real-time low-light image enhancement network model based on a lightweight multi-scale shifted convolutional network, including building a lightweight low-light image enhancement network based on a deep Auto-Encoder architecture, determining the core components of the feature encoder, bottleneck layer, feature decoder and loss function supervision module, and defining the key parameters of the dynamic shifted convolutional kernel DSConv and the multi-scale shifted residual block MSRB and the overall network configuration. S2: Train the real-time low-light image enhancement network model to obtain the optimal model parameters; S3: Build a real-time low-light image enhancement network model adapted to the vehicle embedded platform using the optimal model parameters to enhance the low-light images captured in real time by the vehicle camera.

[0007] Furthermore, the real-time low-light image enhancement network model is built on a deep Auto-Encoder architecture, including a feature encoder, a bottleneck layer, a feature decoder, and a loss function supervision module. The feature encoder progressively downsamples the low-light input image through L downsampled multi-scale shift residual blocks, extracting multi-scale features while reducing resolution. The bottleneck layer consists of a single MSRB, which performs deep processing on the high-dimensional features output by the encoder. The feature decoder gradually restores the image resolution and fuses the features from each stage of the encoder through L upsampled multi-scale shift residual blocks and dense skip connections, preserving detailed information. The loss function supervision module performs multi-stage supervision on the L intermediate feature maps output by the decoder through a composite loss function, optimizing the enhancement effect.

[0008] Furthermore, the dynamic shift convolution kernel DSConv achieves feature extraction through a process of generating offset / residual features via 1×1 pointwise grouped convolution, boundary padding, 3×3 grid sampling, feature aggregation, and gating modulation.

[0009] Furthermore, the mathematical expression of the feature processing flow of DSConv is as follows: The offset / residual feature generation is represented as follows:

[0010] in As an offset feature, For residual characteristics, For the input tensor, For the real number field, B For batch size, C For the number of channels, H and WThese represent the height and width of the feature map, respectively. This is the first 1×1 pointwise grouped convolution; Boundary fill is represented as:

[0011] in For fill operation, The expansion rate of DSConv, with a generation dimension of The filled feature map; Shift sampling is represented as:

[0012] in, Nine shift feature maps are generated. Feature aggregation is represented as:

[0013] Gated modulation is represented as:

[0014]

[0015] in This is the final output of DSConv.

[0016] Furthermore, the multi-scale shift residual block (MSRB) is composed of... The DSConv array is constructed in parallel with different expansion rates.

[0017] Furthermore, step S2 specifically includes the following steps: S21: Initialize network parameters; S22: Divide the low-light image dataset into training and testing sets, and perform preprocessing; S23: Constructing from reconstruction loss Structural similarity loss With multi-level gradient-aware loss The composite loss function consists of:

[0018] in, , and These represent the hyperparameters for each type of loss; S24: Set up a training strategy, train the real-time low-light image enhancement network model using the training set, and test the real-time low-light image enhancement network model using the test set.

[0019] Furthermore, the multi-level gradient-aware loss The gradients in the horizontal and vertical directions are calculated using the Sobel operator, as shown in the following formula:

[0020] in , The decoder outputs at three scales. This is a bilinear interpolation downsampling operator. These are the Sobel gradient operators in the horizontal and vertical directions, respectively.

[0021] Furthermore, step S3 specifically includes the following steps: S31: Input the preprocessed low-light image into the trained real-time low-light image enhancement network model, and output the enhanced image through encoder downsampling, bottleneck layer feature processing, and decoder upsampling; S32: Deploy the model on the vehicle-mounted embedded platform; S33: Evaluate enhanced image quality using full-reference metrics PSNR, SSIM, LPIPS and non-reference metrics NIQE, LOE, DE, EME.

[0022] On the other hand, the present invention provides a low-light image enhancement system based on a lightweight multi-scale shift convolutional network, comprising: Network building module: used to execute step S1, build a real-time low-light image enhancement network model based on a lightweight multi-scale shift convolutional network, including building a lightweight low-light image enhancement network based on a deep Auto-Encoder architecture, determining the core components of the feature encoder, bottleneck layer, feature decoder and loss function supervision module, and defining the key parameters of the dynamic shift convolutional kernel DSConv, the multi-scale shift residual block MSRB and the overall network configuration. Model training module: used to execute step S2, train the real-time low-light image enhancement network model, and obtain the optimal model parameters; Inference Enhancement Module: Used to execute step S3, build a real-time low-light image enhancement network model adapted to the vehicle embedded platform using optimal model parameters, and enhance the low-light images captured in real time by the vehicle camera.

[0023] The beneficial effects of this invention are as follows: (1) Extremely lightweight: The number of parameters is as low as 36 learnable parameters, which is more than 40% lower than the SOTA method. The memory usage is very small, which is suitable for the resource-constrained characteristics of automotive embedded platforms. (2) High efficiency and real-time performance: The computational load is much lower than that of traditional convolution and dilated convolution, and the inference speed is up to 1.72ms / frame, which meets the real-time processing requirements of the vehicle system (frame rate > 500fps). (3) Enhanced quality: On the LoLI-Street (driving scene) dataset, the PSNR reaches 19.27dB and the SSIM reaches 0.92, which exceeds the SOTA method by 4.6dB, effectively solving the problems of color distortion and loss of detail in low light images; (4) Flexible adaptability: It supports parameter configuration of κ=1-5, and can dynamically adjust the model size according to the computing power of the vehicle platform (such as entry-level MCU to high-end AGX Orin), with strong deployment compatibility.

[0024] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 A schematic diagram of the lightweight architecture design of DSConv; Figure 2 This is a schematic diagram of the overall network architecture. Detailed Implementation

[0026] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0027] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0028] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0029] Example 1: To address the limitations of existing deep learning-based low-light image enhancement algorithms, which are constrained by computationally intensive network architectures and struggle to balance real-time processing requirements with image enhancement quality when deployed on automotive embedded platforms, and the technical problems of some lightweight solutions, such as loss of detail and color distortion in complex driving scenarios despite low parameter counts, or insufficient computational efficiency after incorporating new technologies, this invention provides a real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network. This method constructs a convolutional neural network model with a deep Auto-Encoder architecture. It designs a dynamic shifting convolutional kernel (DSConv) with only 12 learnable parameters to replace the traditional convolution. It integrates DSConv kernels with different shift distances to construct multi-scale shifted residual blocks (MSRB) to efficiently extract multi-scale features and expand the receptive field. It introduces a carefully designed residual structure to solve the problem of unstable gradient propagation in lightweight networks. At the same time, it proposes a novel multi-level gradient-aware loss function to enhance training stability and supervision effectiveness. It can also flexibly configure the parameters and number of DSConv kernels to adapt to different processing speed requirements. The minimum number of parameters is only 36 learnable parameters. Finally, it realizes real-time, high-quality enhancement of low-light images on the vehicle embedded platform, which meets the image quality requirements of downstream vision tasks in autonomous driving.

[0030] like Figure 1-2 As shown, this method includes the following steps: S1. Building a Lightweight Multi-Scale Shifted Convolutional Network Model: A lightweight low-light image enhancement network based on a deep Auto-Encoder architecture is built. The core components—feature encoder, bottleneck layer, feature decoder, and loss function supervision module—are determined. Key parameters of the dynamic shifted convolutional kernel (DSConv) and multi-scale shifted residual block (MSRB), along with the overall network configuration, are defined. Specifically, the following steps are included: S11. A real-time low-light image enhancement network based on a lightweight multi-scale shift convolutional network is based on a deep auto-encoder architecture, including a feature encoder, a bottleneck layer, a feature decoder, and a loss function supervision module. The feature encoder uses L downsampled multi-scale shift residual blocks (M... -The low-light input image is progressively downsampled to extract multi-scale features while reducing resolution; the bottleneck layer consists of a single MSRB, which performs depth processing on the high-dimensional features output by the encoder; the feature decoder uses L upsampled multi-scale shift residual blocks (M... + The image resolution is gradually restored and the features of each stage of the encoder are fused together with dense skip connections to preserve detailed information; the loss function supervision module performs multi-stage supervision on the L intermediate feature maps output by the decoder through a composite loss function to optimize the enhancement effect. S12, Core Module Design: (1) Dynamic Shift Convolution Kernel (DSConv): When the number of channels is 3, a single DSConv contains only 12 learnable parameters. It simulates the function of dilated convolution in expanding the receptive field through a process of "1×1 pointwise grouped convolution to generate offset / residual features → boundary padding → 3×3 grid pattern sampling to generate 9 shift feature maps → feature aggregation → gated modulation". (Number of channels) At that time, the computational cost is only 1 / 307 of that of dilated convolution; the mathematical expression of the feature processing flow of DSConv is as follows: (1) Offset / Residual Feature Generation: ,in For the input tensor, This is the first 1×1 pointwise grouped convolution; (2) Boundary filling: The generation dimension is The filled feature map; (3) Shift sampling: ,in, Nine shift feature maps are generated. (4) Feature aggregation: ; (5) Gated modulation: , ,in This is the final output of DSConv.

[0031] (2) Multiscale shifted residual block (MSRB): composed of It consists of several parallel DSConv submodules with different expansion rates, and the total number of parameters is [number missing]. It can efficiently capture multi-scale contextual information. The temporal receptive field is equivalent to an 11×11 dilated convolution; a derived downsampling variant (M) is also possible. - ) and upsampling variant (M + ), where M + It is acceptable to skip connections between features from the previous layer of the decoder and features from the encoder to achieve feature fusion; (3) Network parameter configuration: When At that time, the total number of network parameters was 180 learnable parameters, and the computational cost was 14.036M; when The UltraFast-LieNET Mini reduces the number of parameters to 36 learnable parameters and the computational cost to 2.807M, making it suitable for different computing power requirements.

[0032] S2. Training and Parameter Optimization of a Lightweight Multi-Scale Shifting Convolutional Network Model: The network model is initialized with parameters. A low-light image dataset is prepared and preprocessed. A training strategy and a composite loss function are set. Parameters are updated through error backpropagation to obtain the optimal model parameters. Specifically, the following steps are included: S21. Network parameter initialization: The weights of the 1×1 convolutional layer in DSConv are initialized using the Kaiming Normal algorithm, and the weight scaling factor is set. Initialize the bias terms of all convolutional layers to 0 to ensure stable parameter distribution during the initial training phase and avoid gradient explosion or vanishing. S22. Dataset Preparation: Select four low-light image datasets: LOL, LSRW-HUAWEI, LSRW-NIKON, and LoLI-Street (including the LoLI-Street dataset specifically for driving scenarios), and divide them into training and testing sets in a 9:1 ratio; crop all images from the center into 180×180 pixel blocks, normalize the pixel values ​​to the floating-point range of [0.0, 1.0] using ToTensor transformation, and convert them to PyTorch Tensor format; S23. Training Strategy Settings: The Adam optimizer is used for parameter updates, and training is conducted for 360 epochs; the initial learning rate is set to 0.01, and the learning rate is multiplied by a decay factor every 40 training epochs. Set the batch size to 40; S24. Definition of Composite Loss Function: Constructing a composite loss function from the reconstruction loss... Structural similarity loss With gradient-aware loss Composition of composite loss function .

[0033]

[0034] in, , and These represent the hyperparameters of each loss term. Referring to the design concept of a hybrid L1 and SSIM loss, and The values ​​are 0.975 and 0.025, respectively. The empirical value is taken as 1.

[0035] Gradient-aware loss The gradients in the horizontal and vertical directions are calculated using the Sobel operator, as shown in the following formula: ,in , The decoder outputs at three scales. This is a bilinear interpolation downsampling operator. These are the Sobel gradient operators in the horizontal and vertical directions, respectively.

[0036] S3. Low-light image enhancement inference: An inference model adapted to the vehicle-mounted embedded platform is built using optimal model parameters. Low-light images captured by the vehicle-mounted camera are preprocessed and input into the model for calculation, outputting an enhanced image and verifying the enhancement effect. Specifically, the following steps are included: S31. Inference Calculation: Input the preprocessed low-light image into the trained network model, and output the enhanced image through encoder downsampling (resolution is halved step by step), bottleneck layer feature processing, and decoder upsampling (resolution is restored step by step). S32. Inference Deployment: The model is deployed on an automotive embedded platform (such as Jetson AGX Orin 64GB). The inference time for a 600×400 pixel low-light image is 1.72ms for UltraFast-LieNET Mini and 2.69ms for UltraFast-LieNET Max, which is better than state-of-the-art lightweight methods such as ZeroDCE++ (3.99ms) and meets the requirements for real-time processing. S33. Enhancement effect verification: The results were evaluated using full-reference metrics (PSNR, SSIM, LPIPS) and non-reference metrics (NIQE, LOE, DE, EME). On the LoLI-Street dataset, the PSNR reached 19.27dB and the SSIM reached 0.92, which is 4.6dB (PSNR) higher than the state-of-the-art method.

[0037] Example 2: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0038] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0039] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0040] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily all refer to the same embodiment.

[0041] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0042] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0043] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0044] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0045] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0046] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0047] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network, characterized in that: Includes the following steps: S1: Construct a real-time low-light image enhancement network model based on a lightweight multi-scale shifted convolutional network, including building a lightweight low-light image enhancement network based on a deep Auto-Encoder architecture, determining the core components of the feature encoder, bottleneck layer, feature decoder and loss function supervision module, and defining the key parameters of the dynamic shifted convolutional kernel DSConv and the multi-scale shifted residual block MSRB and the overall network configuration. S2: Train the real-time low-light image enhancement network model to obtain the optimal model parameters; S3: Build a real-time low-light image enhancement network model adapted to the vehicle embedded platform using the optimal model parameters to enhance the low-light images captured in real time by the vehicle camera.

2. The real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network according to claim 1, characterized in that: The real-time low-light image enhancement network model is built on a deep Auto-Encoder architecture, including a feature encoder, a bottleneck layer, a feature decoder, and a loss function supervision module. The feature encoder progressively downsamples the low-light input image through L downsampled multi-scale shift residual blocks, extracting multi-scale features while reducing resolution. The bottleneck layer consists of a single MSRB, which performs deep processing on the high-dimensional features output by the encoder. The feature decoder gradually restores the image resolution and fuses the features from each stage of the encoder through L upsampled multi-scale shift residual blocks and dense skip connections, preserving detailed information. The loss function supervision module performs multi-stage supervision on the L intermediate feature maps output by the decoder through a composite loss function, optimizing the enhancement effect.

3. The real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network according to claim 1, characterized in that: The dynamic shift convolution kernel DSConv extracts features through a process of generating offset / residual features via 1×1 pointwise grouped convolution, boundary padding, 3×3 grid sampling, feature aggregation, and gating modulation.

4. The real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network according to claim 3, characterized in that: The mathematical expression of the feature processing flow of DSConv is as follows: The offset / residual feature generation is represented as follows: in As an offset feature, For residual characteristics, For the input tensor, For the real number field, B For batch size, C For the number of channels, H and W These represent the height and width of the feature map, respectively. This is the first 1×1 pointwise grouped convolution; in For fill operation, For the expansion rate of DSConv, the boundary fill is expressed as: Generation dimension is The filled feature map; Shift sampling is represented as: in, Nine shift feature maps are generated. Feature aggregation is represented as: Gated modulation is represented as: in This is the final output of DSConv.

5. The real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network according to claim 1, characterized in that: The multi-scale shift residual block (MSRB) is composed of The DSConv array is constructed in parallel with different expansion rates.

6. The real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21: Initialize network parameters; S22: Divide the low-light image dataset into training and testing sets, and perform preprocessing; S23: Constructing from reconstruction loss Structural similarity loss With multi-level gradient-aware loss The composite loss function consists of: in, , and These represent the hyperparameters for each type of loss; S24: Set up a training strategy, train the real-time low-light image enhancement network model using the training set, and test the real-time low-light image enhancement network model using the test set.

7. The real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network according to claim 6, characterized in that: The multi-level gradient sensing loss The gradients in the horizontal and vertical directions are calculated using the Sobel operator, as shown in the following formula: in , The decoder outputs at three scales. This is a bilinear interpolation downsampling operator. These are the Sobel gradient operators in the horizontal and vertical directions, respectively.

8. The real-time low-light image enhancement method based on a lightweight multi-scale shift convolutional network according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31: Input the preprocessed low-light image into the trained real-time low-light image enhancement network model, and output the enhanced image through encoder downsampling, bottleneck layer feature processing, and decoder upsampling; S32: Deploy the model on the vehicle-mounted embedded platform; S33: Evaluate enhanced image quality using full-reference metrics PSNR, SSIM, LPIPS and non-reference metrics NIQE, LOE, DE, EME.

9. A low-light image enhancement system based on a lightweight multi-scale shift convolutional network, characterized in that: To implement the method of any one of claims 1-8, comprising: Network building module: used to execute step S1, build a real-time low-light image enhancement network model based on a lightweight multi-scale shift convolutional network, including building a lightweight low-light image enhancement network based on a deep Auto-Encoder architecture, determining the core components of the feature encoder, bottleneck layer, feature decoder and loss function supervision module, and defining the key parameters of the dynamic shift convolutional kernel DSConv, the multi-scale shift residual block MSRB and the overall network configuration. Model training module: used to execute step S2, train the real-time low-light image enhancement network model, and obtain the optimal model parameters; Inference Enhancement Module: Used to execute step S3, build a real-time low-light image enhancement network model adapted to the vehicle embedded platform using optimal model parameters, and enhance the low-light images captured in real time by the vehicle camera.

Images