Low-light image enhancement method and system based on binary neural network

Abstract

The invention discloses a low-light image enhancement method and system based on a binary neural network, and relates to the technical field of low-light image enhancement, and the method comprises the steps: obtaining a full-precision network and a low-light image; a BiLLIE framework is adopted, on the basis of a full-precision network, a binarization module is used for gradually replacing a standard convolution module, and a full-precision dark light enhancement network is constructed; and inputting the low-light image into the full-precision dark light enhancement network to obtain an enhanced low-light image. According to the method, through binarization operation, the memory and the calculation amount are remarkably compressed, and the utilization efficiency of hardware resources is maximized. Through a two-step training strategy, smooth transition from a full-precision model to a binary model is realized. The full-precision weight is kept through activation value binaryzation, and the training stability is ensured and the precision loss is avoided through step-by-step binaryzation of the convolution weight. According to the training strategy, the model can gradually adapt to binarization operation, and the convergence of the training process and the final image enhancement performance are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of low-light image enhancement, and more particularly to a low-light image enhancement method and system based on a binarized neural network. Background Art

[0002] Low-light image enhancement (LLIE) is an important technology in the field of image processing and is widely used in multiple fields such as medical imaging, surveillance systems, and smartphone photography. Its main goal is to improve the quality of images taken in low-light environments, enhancing image clarity, details, and visibility. However, existing low-light image enhancement methods often suffer from high computational overhead, especially on resource-constrained devices (such as edge computing devices, embedded devices, etc.), making it difficult to achieve efficient deployment. Traditional image enhancement methods usually rely on full-precision convolutional neural networks (CNNs), which pose high requirements on the memory and computing capabilities of devices, making it difficult for them to meet the requirements of real-time performance and resource efficiency in practical applications.

[0003] To address this issue, in recent years, binarized neural networks (BNNs), as a method of extreme compression networks, have gradually received extensive attention. BNNs achieve significant memory compression and computational efficiency improvement by binarizing the activation values and weights of neural networks to {-1, +1}. Specifically, BNNs can significantly reduce memory occupancy and increase computational speed without sacrificing performance. For example, BNNs can achieve a 32-fold memory compression ratio and a 64-fold computational acceleration on dedicated hardware. However, existing BNN architectures do not perform well in low-light image enhancement tasks, mainly due to the accuracy loss and color distortion problems introduced during the binarization process. Especially in scenarios where high-quality enhancement effects are required, the visual distortion caused by binarization cannot be ignored.

[0004] Current binarized neural networks (such as BAM, BTM, BBCU, etc.) attempt to solve this problem by improving the binarized convolution design. However, these methods still struggle to find an ideal balance between enhancement effects and computational overhead. For example, although existing BNN models alleviate information loss by introducing strategies such as residual connections, due to complex degradation factors in low-light images (such as low visibility, noise, and color distortion, etc.), existing methods are difficult to effectively suppress these factors, thus affecting the final enhancement effect.

[0005] Therefore, how to propose a low-light image enhancement method and system based on a binarized neural network to achieve efficient low-light image enhancement on resource-limited devices is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a low-light image enhancement method and system based on a binary neural network, which can achieve efficient low-light image enhancement on devices with limited resources. To achieve the above object, the present invention adopts the following technical solutions:

[0007] A low-light image enhancement method based on a binary neural network, comprising:

[0008] Obtaining a full-precision network and a low-light image;

[0009] Adopting the BiLLIE framework, based on the full-precision network, gradually replacing the standard convolution module with a binary module, and dynamically balancing the propagation of binary precision and full-precision information through an adaptive information flow mechanism to construct a full-precision low-light enhancement network;

[0010] Inputting the low-light image into the full-precision low-light enhancement network to obtain an enhanced low-light image.

[0011] Optionally, the standard convolution module includes a full-precision channel shuffle module, a full-precision basic convolution block, and a full-precision up / down sampling module, and the binary module includes a binary channel fusion module, a BiLLIE basic convolution unit, and a binary up / down sampling module.

[0012] Optionally, the step-by-step replacement of the standard convolution module with the binary module includes replacing the full-precision channel shuffle module with the binary channel fusion module, replacing the full-precision basic convolution block with the BiLLIE basic convolution unit, and replacing the full-precision up / down sampling module with the binary up / down sampling module.

[0013] Optionally, it further includes that the first convolution block and the last convolution block in the binary module are not replaced by binary, and a two-step training strategy is adopted to systematically replace the convolution operator through four configurations.

[0014] Optionally, the four configurations include the network structure of the conv-ReLU sequence, the Pad-conv-ReLU sequence, a single convolution with a residual connection, and an independent convolution in the LLIE model.

[0015] Optionally, the full-precision low-light enhancement network includes:

[0016] S1: Binarizing the activation value X of each module image into X f while keeping the full-precision weight W b ; f ;

[0017] S2: On the basis of S1, binarizing the convolution weight in the network into W b and making the network gradually adapt to the binary operation, changing from full-precision to binary.

[0018] Optionally, the specific structure of the full-precision low-light enhancement network includes:

[0019] S11: Let and X b represent the full-precision and binary input feature maps respectively, and W b represent the full-precision and binary weight convolution kernels respectively. The binarization operation of the excitation is defined as follows:

[0020] where, is the full-precision input feature value of the l-th layer, is the output feature value after binarization, taking values of +1 or -1, and α l is the learned channel-level threshold adjustment parameter;

[0021] S22: The binary weight is defined as follows:

[0022]

[0023]

[0024] where, is the full-precision weight matrix of the c-th channel of the l-th layer, C out is the number of output channels, K is the size of the convolution kernel, is the binarized weight value of the l-th layer, taking values of +1 or -1, is the full-precision weight value of the l-th layer. Among them, the scaling factor s c,l ∈S l , introduced by XNOR-Net, aims to approximate the full-precision weight

[0025] Optionally, the activation function of the full-precision low-light enhancement network is defined as:

[0026]

[0027] where, x c,l ∈R H×W×C represents the input feature map of RPReLU, and γ c,l , ζ c,l , β c,l are learnable parameters, c represents the channel index, and l represents the layer index.

[0028] Optionally, the dynamic balance of binary precision and full-precision information propagation through the adaptive information flow mechanism includes the AIF module:

[0029]

[0030] Among them, represents the adjusted residual, is the channel-specific self-adjusting weight vector, used to selectively retain or suppress the feature channels in is the channel-level adaptive bias term, is a real vector space of dimension C_in, is the feature map of the previous layer, with dimension R H×W×Cin ;

[0031] Through the AIF module, the complete binary convolutional block of the l-th layer is represented as:

[0032]

[0033] Among them, ψ(·) represents the non-linear activation function, represents the efficient binary convolution operation implemented through XNOR and bitcount operations, S l contains the scaling factor for approximating full-precision convolution through broadcasting.

[0034] Optionally, a low-light image enhancement system based on a binary neural network includes:

[0035] An acquisition module: used to acquire a full-precision network and a low-light image;

[0036] A construction module: used to adopt the BiLLIE framework, based on the full-precision network, gradually replace the standard convolution module with a binary module, and dynamically balance the propagation of binary precision and full-precision information through an adaptive information flow mechanism to construct a full-precision low-light enhancement network;

[0037] An enhancement module: used to input the low-light image into the full-precision low-light enhancement network to obtain an enhanced low-light image.

[0038] From the above technical solutions, it can be seen that compared with the prior art, the present invention discloses a low-light image enhancement method and system based on a binary neural network, having the following beneficial effects:

[0039] The present invention proposes a low-light image enhancement method based on a binary neural network, including: obtaining a full-precision network and a low-light image; adopting the BiLLIE framework, and on the basis of the full-precision network, gradually replacing the standard convolution module with a binary module, and dynamically balancing the propagation of binary precision and full-precision information through an adaptive information flow mechanism to construct a full-precision low-light enhancement network; inputting the low-light image into the full-precision low-light enhancement network to obtain an enhanced low-light image. The present invention: (1) Through binary operation, the BiLLIE framework significantly compresses memory and computational amount, and maximizes the utilization efficiency of hardware resources. In the test, the computational and storage costs of the BiLLiE framework are reduced by more than about 90% compared with the full-precision model, so that the low-light image enhancement technology can be widely applied in resource-constrained environments; (2) While controlling degradation factors such as noise and color distortion in low-light images, the BiLLiE framework maintains the details and brightness of the image through an adaptive information flow mechanism, and reduces visual distortion caused by binaryization. Experimental results show that the BiLLiE framework significantly outperforms other BNN models in terms of PSNR and SSIM metrics, and the enhancement quality is close to that of the full-precision model; (3) Through a two-step training strategy, the BiLLiE framework realizes a smooth transition from the full-precision model to the binary model. The first step maintains the full-precision weight through activation value binaryization, and the second step gradually binaryizes the convolutional weight to ensure training stability and avoid accuracy loss. This training strategy enables the model to gradually adapt to binary operation, greatly improving the convergence of the training process and the final image enhancement performance; (4) The BiLLiE framework proposed by the present invention can be applied to a variety of LLIE benchmark models, including a series of convolution neural network-based low-light enhancement models such as LEDNet, PairLIE, and RUAS, which cover both lightweight and computationally burdensome models, and superior performance is demonstrated on these methods. Compared with other binary methods, the BiLLiE framework significantly improves the image enhancement effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0041] Figure 1 It is a schematic flowchart of a low-light image enhancement method based on a binary neural network provided by the present invention.

[0042] Figure 2 It is a schematic structural diagram of a full-precision low-light enhancement network provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] An embodiment of the present invention discloses a low-light image enhancement method based on a binary neural network, including:

[0045] Obtaining a full-precision network and a low-light image;

[0046] Adopting the BiLLIE framework, based on the full-precision network, gradually replacing the standard convolution module with a binary module, and dynamically balancing the propagation of binary precision and full-precision information through an adaptive information flow mechanism to construct a full-precision low-light enhancement network;

[0047] Inputting the low-light image into the full-precision low-light enhancement network to obtain an enhanced low-light image.

[0048] Further, the standard convolution module includes a full-precision channel shuffling module, a full-precision basic convolution block, and a full-precision up / down sampling module, and the binary module includes a binary channel fusion module, a BiLLIE basic convolution unit, and a binary up / down sampling module.

[0049] Further, the step of gradually replacing the standard convolution module with a binary module includes replacing the full-precision channel shuffling module with a binary channel fusion module, replacing the full-precision basic convolution block with a BiLLIE basic convolution unit, and replacing the full-precision up / down sampling module with a binary up / down sampling module.

[0050] Further, it also includes that the first convolution block and the last convolution block in the binary module are not replaced by binary, and a two-step training strategy is adopted to systematically replace the convolution operator through four configurations.

[0051] Further, the four configurations include the network structure of the conv-ReLU sequence, the Pad-conv-ReLU sequence, a single convolution with a residual connection, and an independent convolution in the LLIE model.

[0052] Further, the full-precision low-light enhancement network includes:

[0053] S1: Binarize the activation value X of each module image f to X b , and keep the full-precision weight W f ;

[0054] S2: Based on S1, binarize the convolutional weights in the network into W b , and let the network gradually adapt to the binary operation, changing from full precision to binary.

[0055] Furthermore, the specific structure of the full-precision low-light enhancement network includes:

[0056] S11: Let and X b represent the full-precision and binary input feature maps respectively, and W b represent the full-precision and binary weight convolutional kernels respectively. The binarization operation of the excitation is defined as follows:

[0057] where, is the full-precision input feature value of the l-th layer, is the output feature value after binarization, taking values of +1 or -1, and α l is the learned channel-level threshold adjustment parameter;

[0058] S22: The binary weight is defined as follows:

[0059]

[0060]

[0061] where, is the full-precision weight matrix of the c-th channel of the l-th layer, C out is the number of output channels, K is the size of the convolutional kernel, is the weight value after binarization of the l-th layer, taking values of +1 or -1, is the full-precision weight value of the l-th layer. Among them, the scaling factor s c,l ∈S l , introduced by XNOR-Net, aims to approximate the full-precision weight

[0062] Furthermore, the activation function of the full-precision low-light enhancement network is defined as:

[0063]

[0064] where, x c,l ∈R H×W×C represents the input feature map of RPReLU, and γ c,l , ζ c,l , β c,l are learnable parameters, c represents the channel index, and l represents the layer index.

[0065] Furthermore, the dynamic balancing of the propagation of binary precision and full precision information through the adaptive information flow mechanism includes an AIF module:

[0066]

[0067] Among them, represents the adjusted residual, is the channel-specific self-adjusting weight vector for selectively retaining or suppressing the feature channels in is the channel-level adaptive bias term, is a real vector space of dimension Cin, is the feature map of the previous layer, with dimension R H×W×Cin ;

[0068] Through the AIF module, the complete binary convolution block of the l-th layer is expressed as:

[0069]

[0070] Among them, ψ(·) represents the non-linear activation function, represents the efficient binary convolution operation implemented through XNOR and bitcount operations, S l contains the scaling factor for approximating the full precision convolution through broadcasting.

[0071] In a specific embodiment, a low-light image enhancement system based on a binary neural network includes:

[0072] An acquisition module: used to acquire the full precision network and the low-light image;

[0073] A construction module: used to adopt the BiLLIE framework, and on the basis of the full precision network, gradually replace the standard convolution module with a binary module, and dynamically balance the propagation of binary precision and full precision information through the adaptive information flow mechanism to construct a full precision low-light enhancement network;

[0074] An enhancement module: used to input the low-light image into the full precision low-light enhancement network to obtain the enhanced low-light image.

[0075] In a specific embodiment, a low-light image enhancement method based on a binary neural network, as Figure 1 shown, (a) the binary channel fusion module for channel dimension adjustment; (b) the BiLLIE basic convolution unit; (c) the binary up / down sampling module for spatial resolution adjustment; (d) the internal structure of the adaptive information flow module for dynamically selecting full precision information; (e) the binary convolution accelerated by using the XNOR-Bitcount operation, and the specific steps are as follows:

[0076] S1. Obtain a low-light image and a full-precision low-light enhancement network: First, obtain a low-light image I low and input it into the full-precision low-light enhancement network. The network adopts the BiLLIE framework and, based on the full-precision network, gradually replaces it with binarized modules.

[0077] S2. Replace binarized modules: In the full-precision model, according to Figure 2 , where (a)-(b) are two-step training strategies, and the convolutional operators are systematically replaced through four configurations, including: (c) the network structure of the conv-ReLU sequence, (d) the Pad-conv-ReLU sequence, (e) a single convolution with a residual connection, and (f) an independent convolution in the LLIE model. Gradually replace the standard convolution modules with binarized modules. These binarized modules include binarized channel fusion modules, BiLLIE basic convolutional units, and binarized up / downsampling modules. The first and last convolutional blocks in the network should not be replaced with binarized modules.

[0078] S3. First-step training: In the first step of training, binarize the activation value X f of each module image into X b , and keep the full-precision weight W f . This step helps the network gradually adapt to the binarization operation while retaining the quality of image enhancement.

[0079] S4. Second-step training: Based on the first step, binarize the convolutional weights in the network into W b . Through this process, the network gradually adapts to the binary operation and completes the transformation from full-precision to binarized.

[0080] S5. Obtain the binarized low-light enhancement network, which can output a high-quality picture I enhanced while saving a large amount of computational burden.

[0081] The present invention proposes a framework called BiLLIE (Binarized Low Light Image Enhancement), which introduces the principle of binarized neural networks into low-light image enhancement, aiming to significantly reduce the computational complexity through the binarization of activation values and weights. By restricting the activation values and weights to 1-bit representations, binarized neural networks can achieve up to 32 times of memory compression and 64 times of reduction in computational volume, enabling the low-light enhancement technology to be efficiently applied to restricted hardware environments.

[0082] Furthermore, to address the issues of accuracy loss and color distortion faced by traditional binary neural networks in low-light enhancement tasks, the present invention proposes an Adaptive Information Flow (AIF) mechanism. This mechanism dynamically balances the propagation of binary precision and full-precision information, effectively solving the problem of data discontinuity unique to low-light enhancement. In addition, the present invention also proposes a specialized binary training strategy. By gradually adapting to binary operations, the network can maximize the retention of image enhancement quality while maintaining efficient computation. Through comprehensive experimental verification, the application of the present invention in multiple low-light enhancement networks shows that the BiLLiE framework can significantly reduce memory and computational requirements while maintaining enhancement quality comparable to that of full-precision models, making it suitable for practical application scenarios that require high efficiency and low power consumption.

[0083] In the specific implementation, the present invention proposes a binary neural network architecture applicable to low-light image enhancement tasks, called BiLLIE (Binarized Low Light Enhancer). This architecture achieves a balance between computational efficiency and enhancement effect through the adaptive information flow mechanism and an improved binary training strategy. The basic unit design of the BiLLiE framework is as Figure 1 shown, including the dynamic fusion of full-precision feature maps and binary feature maps. To optimize the information flow during the binarization process, the present invention introduces an adaptive scaling and fusion mechanism.

[0084] 1) Basic unit design

[0085] As Figure 1 shown, let and X b represent the full-precision and binary input feature maps respectively, and and W b represent the full-precision and binary weight convolution kernels respectively. To binarize the full-precision excitation, the binarization operation of the excitation is defined as follows:

[0086] where, is the full-precision input feature value (original floating-point value) of the l-th layer, is the output feature value after binarization, taking values of +1 or -1, and α l is the learned channel-level threshold adjustment parameter, used to dynamically adjust the activation threshold and suppress degradation factors (such as noise, color distortion) in low-light images. Among them, represents the binary excitation value, and α l represents the learnable parameter, used to adjust the activation threshold of the sign function specific to the l-th layer in the neural network.

[0087] The binary weight is defined as follows:

[0088]

[0089]

[0090] Among them, is the full-precision weight matrix of the c-th channel in the l-th layer, and C out is the number of output channels, and K is the size of the convolution kernel. is the binarized weight value of the l-th layer, taking values of +1 or -1. is the full-precision weight value (original floating-point value) of the l-th layer, where the scaling factor s c,l ∈S l is introduced by XNOR-Net to approximate the full-precision weights.

[0091] The activation function is defined as:

[0092]

[0093] where x c,l ∈R H×W×C represents the input feature map of RPReLU, and γ c,l , ζ c,l , β c,l are learnable parameters, c represents the channel index, and l represents the layer index.

[0094] 2) Adaptive Information Flow

[0095] Since Bi-real net emphasizes the importance of full-precision information flow in BNN, subsequent methods have also continuously adopted this paradigm. However, in the context of low-light enhancement, continuous full-precision information flow may hinder data decomposition and lead to color inconsistency. As Figure 1 shown in (d), to address the limitations of continuous full-precision information flow in low-light enhancement tasks, an Adaptive Information Flow (AIF) mechanism is proposed, which can dynamically regulate the propagation of information:

[0096]

[0097] where represents the adjusted residual, is the channel-specific self-adjusting weight vector used to selectively retain or suppress the feature channels in is the channel-level adaptive bias term, is a real vector space of dimension Cin, is the feature map of the previous layer, with dimension RH×W×Cin 。

[0098] Through the AIF module, as shown in (b) of Figure 2 , the complete binary convolutional block of the l-th layer is expressed as:

[0099]

[0100] where ψ(·) represents the non-linear activation function, represents the efficient binary convolution operation implemented through XNOR and bitcount operations, and S l contains the scaling factor for approximating the full-precision convolution through broadcasting.

[0101] The binary convolution replaces the traditional convolution operation through efficient bitcount and XNOR operations, while significantly accelerating the calculation. This enables the binary convolution to effectively approximate the full-precision convolution operation Y f = W f * X f , where Y f , W f , X f are the output feature map of the full-precision convolution, the full-precision convolution kernel, and the full-precision input feature map.

[0102] Furthermore, it is also crucial to maintain the adaptive information flow in other components of the network. As shown in Figure 1 , a dedicated binary module is designed for channel fusion and up / downsampling operations. The channel fusion module ( Figure 1 in (a)) includes a 1×1 full-precision convolution in its residual path as a dedicated information channel at the critical channel dimension change. Similarly, the refined up / downsampling module, such as Figure 1 in (c), is also designed based on the same principle because the change in spatial resolution may lead to significant feature degradation in the binary network. Through this comprehensive design, an unobstructed network information flow is established at different operation scales and feature dimensions, thereby improving the overall performance of the binary network.

[0103] 3) Low-light enhancement binarization process

[0104] To improve the binarization adaptability of the network, the present invention proposes a low-light enhancement binarization process, aiming to smoothly convert the full-precision model into a binarized model and maintain the stability and enhancement effect during the training process.

[0105] 1. Activation value binarization: First, binarize the activation values of the network while keeping the full-precision weights unchanged to reduce the gradient instability in the initial stage of training. Specifically, the convolution operation is implemented as where X bis the binarized activation value, and W f is the full-precision weight.

[0106] 2. Weight binarization: Based on the binarization of the activation value, the convolutional weights are gradually binarized to form a binarized convolutional operation where W b is the binarized convolutional weight. In this process, all full-precision convolutional modules are replaced with the binarized convolutional blocks proposed by the present invention.

[0107] Specifically, as shown in (c)-(f) of Figure 2 , the convolutional module is gradually replaced with the BiLLiE binarized convolutional block through four configurations to maintain stability and enhancement effect during the training process. During this process, the loss function of the network is consistent with that of the full-precision model, ensuring the gradual maintenance of accuracy during the training process. And through gradual binarization, the accuracy loss is avoided, the computational efficiency is optimized, and at the same time, a high image enhancement quality is maintained. The binarization training strategy of the present invention greatly simplifies the binarization process of the low-light image enhancement network, while improving the training stability and convergence speed.

[0108] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description of the method part.

[0109] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A low-light image enhancement method based on a binarization neural network, characterized in that: include: Get full-precision web and low-light images; Adopting the BiLLIE framework, on the basis of the full-precision network, the standard convolutional modules are gradually replaced with binarization modules, and the propagation of binary precision and full-precision information is dynamically balanced through the adaptive information flow mechanism to build a full-precision dark light enhancement network; The low-light image is input into the full-precision dark-light enhancement network to obtain an enhanced low-light image.

2. The low-light image enhancement method based on a binarization neural network according to claim 1, characterized in that: The standard convolution module includes a full-precision channel confusion module, a full-precision basic convolution block and a full-precision up / down sampling module, and the binarization module includes a binarization channel fusion module, a BiLLIE basic convolution unit and a binary up / down sampling module.

3. The low-light image enhancement method based on a binarization neural network according to claim 2, characterized in that: The step of gradually replacing the standard convolution module with the binarization module includes replacing the full-precision channel confusion module with a binarization channel fusion module, replacing the full-precision basic convolution block with a BiLLIE basic convolution unit, and replacing the full-precision up / down sampling module with a binary up / down sampling module.

4. The low-light image enhancement method based on a binarization neural network according to claim 3, characterized in that: It also includes that the first convolution block and the last convolution block in the binarization module are not binarized, a two-step training strategy is adopted, and the convolution operators are systematically replaced through four configurations.

5. The low-light image enhancement method based on a binarization neural network according to claim 4, characterized in that: The four configurations include a network structure of conv-ReLU sequence, Pad-conv-ReLU sequence, a single convolution with residual connection, and independent convolution in the LLIE model.

6. The low-light image enhancement method based on a binarization neural network according to claim 1, characterized in that: The full-precision dark light enhancement network includes: S1: The activation value X of each module image f Binarize to X b , keeping the full precision weight W f ; S2: Based on S1, the convolution weights in the network are binarized into W b , so that the network gradually adapts to binary operations and transforms from full precision to binary.

7. The low-light image enhancement method based on a binarization neural network according to claim 6, characterized in that: The specific structure of the full-precision dark light enhancement network includes: S11: Set and X b denote the full precision and binary input feature maps respectively, and W b Represent the full-precision and binary weight convolution kernels respectively, and the binarization operation of the excitation is defined as follows: in, is the full-precision input feature value for the lth layer, is the output feature value after binarization, which takes the value of +1 or -1, α l Tuning parameters for learned channel-level thresholds; S22: Binary weights are defined as follows: in, is the full-precision weight matrix of the cth channel in the lth layer, C out is the number of output channels, K is the size of the convolution kernel, is the binarized weight value of the lth layer, which takes the value of +1 or -1. is the full-precision weight value of the lth layer, where the scaling factor s c,l ∈S l , introduced by XNOR-Net, aims to approximate the full-precision weights 8. The low-light image enhancement method based on a binarization neural network according to claim 7, characterized in that: The activation function of the full-precision dark light enhancement network is defined as: Among them, x c,l ∈R H×W×C represents the input feature map of RPReLU, γ c,l , c,l , β c,l are learnable parameters, c represents the channel index, and l represents the layer index.

9. The low-light image enhancement method based on a binarization neural network according to claim 1, characterized in that: The propagation of information dynamically balancing binary precision and full precision through the adaptive information flow mechanism includes the AIF module: in, represents the adjusted residual, is a channel-specific self-adjusting weight vector used to selectively retain or suppress The feature channels in is the channel-level adaptive bias term, is a real vector space with dimension Cin, is the feature map of the previous layer, with a dimension of R H×W×Cin ; Through the AIF module, the complete binary convolution block of layer l is expressed as: Among them, ψ(·) represents the nonlinear activation function, represents the efficient binary convolution operation implemented by XNOR and bitcount operations, S l Contains the scaling factor used to approximate full-precision convolution via broadcasting.

10. A low-light image enhancement system based on a binarization neural network, characterized in that: include: Acquisition module: used to acquire full-precision network and low-light images; Building module: It is used to adopt the BiLLIE framework, gradually replace the standard convolution module with the binarization module on the basis of the full-precision network, dynamically balance the propagation of binary precision and full-precision information through the adaptive information flow mechanism, and build a full-precision dark light enhancement network; Enhancement module: used to input the low-light image into the full-precision dark-light enhancement network to obtain an enhanced low-light image.

Images