Remote gating image enhancement method based on improved JCRNet

Abstract

The invention discloses a long-distance gating image enhancement method based on an improved JCRNet, and the method improves a gating image enhancement algorithm on the basis of the JCRNet, and employs the algorithm to enhance a long-distance gating image, so as to improve the quality of the gating image in a complex environment. Firstly, an enhanced self-calibration convolutional ESCConv network is adopted, edge and high-frequency information is enhanced through a maximum pooling branch, a receptive field is expanded in combination with a multi-branch structure so as to relieve the problems of high-frequency detail loss and edge blur, meanwhile, a LeakReLU activation function is added, the nonlinear mapping capability is enhanced, and the adaptability of a model to complex illumination is improved. Secondly, a feature enhancement and fusion module of the gated image is designed, and by dynamically balancing global and local features, contrast perception is enhanced, local saliency of the gated image is improved, noise interference is effectively suppressed, and visual quality is enhanced. Finally, an ADNet denoising module is added, an attention guiding strategy is adopted to adapt to non-uniform noise distribution in a complex environment, and it is ensured that the enhancement result is stable and reliable.

Description

Technical Field

[0001] This invention belongs to the field of machine vision technology, specifically a long-range gating image enhancement method based on an improved JCRNet. Background Technology

[0002] With the rapid development of intelligent sensing technology, computer vision has been widely applied in fields such as autonomous driving, security monitoring, and target recognition, placing higher demands on image quality in complex environments. Especially in long-distance imaging tasks, improving image clarity, contrast, and detail has become a key challenge. Laser gating imaging, as an active imaging method with precise depth sensing capabilities, effectively suppresses background scattering interference by receiving target reflections within a preset time window, demonstrating significant advantages in dealing with adverse weather and varying lighting conditions. Its basic principle is to use short-pulse lasers to illuminate the target object and filter effective information based on the time delay of the echo signal in order to achieve precise capture of the target area. .

[0003] However, in long-distance imaging, laser signal energy attenuation is significant. Combined with environmental noise, unstable sensor response, and multipath scattering effects, gated images often suffer from reduced contrast, loss of detail, and increased speckle noise. Particularly in backlit, highly reflective, or shadowed areas, local brightness is prone to overexposure or information loss, severely impacting overall visual quality. These image degradation phenomena not only weaken visual perception but also directly restrict the extraction and utilization of effective features in subsequent visual tasks. Therefore, optimizing the quality of gated images under long-distance imaging conditions to improve their performance in complex environments has become a crucial issue urgently needing to be addressed in the field of visual enhancement.

[0004] In the field of low-light image enhancement, traditional algorithms have achieved certain results, but still have many limitations in handling varying lighting conditions. With the rapid rise of deep learning, deep network-based methods have shown strong potential for image enhancement. Existing methods can be broadly divided into Retinex-based methods and Convolutional Neural Network (CNN)-based methods. Based on Retinex theory, Guo et al. A structure-aware smoothing model was proposed to improve image quality in low-light environments; Li et al. By introducing a noise map and combining it with L1 regularization constraints, the contrast of low-light images was improved. (Li et al.) Furthermore, based on the CNN framework, a low-light image synthesis method is proposed, which significantly improves both brightness and contrast. Hao et al. A semi-decoupled Retinex decomposition model was designed, which can effectively suppress image degradation under different noise levels. (Wang et al.) They then optimized the detail rendering effect of low-light images by introducing back-projection blocks and feature aggregation blocks through a deep brightening network; Hossein et al. Based on an autoencoder structure combined with attention mechanisms and skip connections, the robustness of target detection in low-light environments has been further improved. These methods have made significant progress in the field of low-light image enhancement, but for the complex environment of laser-gated imaging, further exploration is needed to find comprehensive optimization methods that can balance noise suppression, detail preservation, and contrast enhancement.

[0005] Gated image enhancement, as an important preprocessing stage, not only improves visual representation but also plays a crucial role in downstream detection tasks. In the field of low-light image enhancement, although there are many improvement methods, most of these methods consider the overall brightness of the image, ignoring local details, local brightness ranges, and edge information loss, leading to problems such as global exposure imbalance in the output gated image. Summary of the Invention

[0006] To address the aforementioned issues, this paper proposes an end-to-end enhancement method for gated images. This method comprehensively considers multiple key factors, including feature extraction, illumination equalization, texture enhancement, and noise suppression, to improve image quality from multiple levels, thereby enhancing the adaptability and robustness of gated imaging in complex environments. The specific details of this invention are as follows.

[0007] A long-range gated image enhancement method based on an improved JCRNet is proposed, which improves the Joint Correction and Refinement Network (JCRNet) and uses the improved gated image enhancement algorithm to enhance long-range gated images.

[0008] The method for improving JCRNet to obtain the long-range gating image enhancement method of the present invention includes the following steps:

[0009] 1) In the feature extraction stage of JCRNet, the residual block is replaced by the enhanced self-calibrating convolution ESCConv module. The edge, texture and high frequency information are enhanced by introducing the max pooling branch. The receptive field is expanded by the multi-branch structure to adapt to distant targets. At the same time, LeakReLU activation is used to improve the nonlinear mapping capability.

[0010] 2) In the joint refinement stage of JCRNet, the Adaptive Single-channel Feature Refinement (ASFR) module is used to process the features extracted in step 1), dynamically balancing global and local features and optimizing contrast and texture details.

[0011] In the feature extraction stage of step 1), for a given input image X with C input channels:

[0012] First, X is divided into X1 and X2, with a channel number of C / 2; at the same time, X is processed by MaxPooling to form a high-frequency component X3;

[0013] Then, X1 is transformed by convolution kernel K1 to obtain the main branch feature Y1; X2 is fed into the ESCConv module and processed by three convolution kernels K2, K3 and K4 with different receptive fields in sequence to finally obtain the output Y2; X3 is first compressed by 1×1 convolution K5, and then restored to its original size by Upsample to obtain Y3; K1 and K4 both use the LeakReLU activation function;

[0014] Finally, Y1, Y2 and Y3 are concatenated to obtain the final output feature map Y;

[0015] In the joint refinement stage of step 2), the feature enhancement and fusion module for the gated image consists of an adaptive feature enhancement module (ASFE) and an adaptive spatial channel fusion module (ASCF); for Y:

[0016] 2.1) In Adaptive Feature Enhancement (ASFE), Y is first processed by multiple convolutional branches of different scales, and the processing results are concatenated along the channel dimension to form a fused feature map; then, the adaptive channel enhancement (ACE) mechanism is used to weight the feature map to obtain Y'.

[0017] 2.2) The Adaptive Spatial-Channel Fusion Module (ASCF) includes a channel attention (CA) branch and a spatial attention (SA) branch; these branches act on the local feature Y1' and global feature Y2' output by the adaptive feature enhancement module, respectively. Specific steps include:

[0018] a. In the channel attention CA path, for Y1', local max pooling (LMP) and adaptive local pooling (ALP) are used to extract local features; for Y2', global max pooling (GMP) and global average pooling (GAP) are used to capture overall features. The four pooling results are concatenated into a fused feature description vector Z, which is then nonlinearly transformed by a multilayer perceptron (MLP) and normalized using a sigmoid function to obtain the channel attention weight mappings corresponding to Y1' and Y2', respectively. and ;

[0019] b. In spatial attention (SA), Y1' first extracts preliminary features through 1×1 convolution, and then applies region adaptive pooling (RAP) to generate a spatial weight matrix K; Y2' extracts global spatial features through global average pooling (GAP) combined with a convolution module to form a weight vector V.

[0020] Subsequently, K and V are multiplied by matrix, and then normalized using a combination of Softmax and Sigmoid to obtain the final spatial attention weights. and ; and Corresponding to Y1' and Y2' respectively;

[0021] Finally, the final fused features are obtained by performing element-wise multiplications on the two branches Y1' and Y2' with their corresponding channel and spatial weights, and then summing the results in a weighted manner. :

[0022] .

[0024] The input image X is processed by the IAS lighting adjustment module of JCRNet and then added to the result processed in step 2). The result is then further processed by the attention-guided denoising network ADNet.

[0025] The processing of X2 in the ESCConv module includes:

[0026] 1.2.1) The results of X2 being processed sequentially by AvgPooling, K3, and Upsample are added to X2;

[0027] 1.2.2) X2 is processed by K2;

[0028] 1.2.3) After activation processing, the result of 1.2.1) is multiplied by the result of 1.2.2), and then processed by K4 to obtain Y2.

[0029] In step 2.1), the ASFE has four convolutional branches: the first branch has a 1×1 convolution; the second branch has 1×1, 3×1, 1×3 and 3×3 convolutions in sequence; the third branch has 1×1, 1×3, 3×1 and 3×3 convolutions in sequence; and the fourth branch has 1×1 and 3×3 convolutions in sequence, which serve as residual paths. The results of Y processed by the second, third and fourth branches are concatenated and then added to the result of the first branch to obtain the fused feature map.

[0030] In the feature extraction stage, the ResCAB module is used to further process the output of the ESCConv module. The processing result is then processed by the encoder-decoder module and then by the self-supervised module SSB to align the preliminary features with the features output by ESCConv and generate a pseudo-supervised signal.

[0031] The encoder and decoder in the encoder-decoder module are connected by a jumper.

[0032] The long-range gated image enhancement method (JGCRNet) based on the improved Joint Correcting and Refinement Network (JCRNet) of this invention mainly comprises the following four core modules:

[0033] (1) To address the issues of high-frequency detail loss and edge blurring in gated images, an extended spatial-channel convolution module ESCConv is introduced as a feature extraction method. Max pooling is introduced to enhance edge, texture, and high-frequency information, and the receptive field is expanded through a multi-branch structure to adapt to distant targets. LeakReLU activation is also used to improve nonlinear mapping capabilities.

[0034] (2) To address the problem that global statistical information is susceptible to noise interference and local saliency is insufficient, an adaptive feature enhancement and fusion module ASFR is designed to dynamically balance global and local features and optimize contrast and texture details.

[0035] (3) Dynamic illumination compensation is performed through the IAS illumination adjustment module, which significantly improves the problems of uneven illumination and loss of detail in dark areas. The illumination adjustment part adopts the IAS illumination adjustment module of the standard JCRNet.

[0036] (4) An attention-guided denoising network ADNet is introduced. Through the division of labor and cooperation of the SB, FEB, AB and RB modules, comprehensive noise suppression and effective enhancement of target information are achieved. Adaptive noise suppression is performed to improve the detail fidelity of the enhanced image. Attached Figure Description

[0037] Figure 1 This diagram shows the structure of the improved algorithm JGCRNet.

[0038] Figure 2 This represents the structure diagram of ESCConv;

[0039] Figure 3 This represents the ASFE structure diagram;

[0040] Figure 4This represents the ASCF structure diagram;

[0041] Figure 5 Represents the ADNet architecture diagram;

[0042] Figure 6 This represents a partially gated dataset;

[0043] Figure 7 This represents a visual comparison of the gated dataset across multiple augmentation algorithms;

[0044] Figure 8 This represents a visual comparison of the LOL dataset across various augmentation algorithms. Detailed Implementation

[0045] Overview

[0046] Laser range-gated systems can achieve long-distance imaging, but the quality of the gated image remains significantly limited in complex scenes due to environmental noise, uneven illumination, and loss of high-frequency details. Therefore, this invention improves the gated image enhancement algorithm based on JCRNet to enhance the quality of gated images in complex environments.

[0047] First, an extended spatial-channel convolution module ESCConv is used to enhance edge and high-frequency information through max pooling branches, and the receptive field is expanded by combining a multi-branch structure to alleviate the problems of high-frequency detail loss and edge blurring. At the same time, the LeakReLU activation function is introduced to enhance nonlinear mapping ability and improve the model's adaptability to complex lighting.

[0048] Secondly, the Feature Enhancement and Fusion (ASFR) module for gated images is introduced. By dynamically balancing global and local features, it enhances contrast perception, improves the local saliency of gated images, effectively suppresses noise interference, and enhances visual quality.

[0049] Finally, an attention-guided denoising convolutional neural network (ADNet) is added, employing a region-adaptive noise suppression strategy to adapt to the non-uniform noise distribution in complex environments, ensuring stable and reliable enhancement results.

[0050] Experimental results show that PSNR is improved by 0.6579 dB and SSIM is improved by 1.32% on the self-built gated dataset, and PSNR is improved by 0.7014 dB and SSIM is improved by 1.46% on the LOL dataset, demonstrating strong robustness and generalization ability.

[0051] The present invention will be further described below with reference to the accompanying drawings.

[0052] 1. Algorithm of this invention

[0053] The improved gated image enhancement algorithm JGCRNet constructs a staged coordinated enhancement framework, which includes: a feature extraction part, a gated image feature enhancement and fusion part, an illumination adjustment part, and a denoising part. Among them, the feature extraction part, the gated image feature enhancement and fusion part, and the denoising part are improved, while the illumination adjustment part adopts the IAS illumination adjustment module of the standard JCRNet.

[0054] The long-range gating image enhancement algorithm framework of this invention is as follows: Figure 1 As shown, this addresses the problem of enhancing gated images in complex scenes.

[0055] The invention will now be further explained in conjunction with the enhancement processing of long-distance gated images.

[0056] 1.1 Improved Feature Extraction Stage EFES

[0057] To address the urgent need for capturing high-frequency details and achieving global semantic understanding in gated images under low-light and high-noise environments, this invention improves upon the traditional Feature Extraction Module (FES) by proposing the Enhanced Feature Extraction Stage (EFES) structure. This structure introduces an extended spatial-channel convolution module, ESCConv, which enhances the perception of high-frequency features through a multi-branch design and integrates the LeakReLU activation function to improve non-linear representation. This enables deep joint mining of local edge features and overall structural information, effectively enhancing the richness and robustness of feature extraction.

[0058] While self-calibrating convolutions excel in enhancing feature extraction capabilities, their design is primarily geared towards natural images, and they still have certain limitations in practical applications for gated images. For example, self-calibrating convolutions are insufficiently adaptable to the uneven illumination problem commonly found in gated images, exhibiting limited edge detection capabilities and easily leading to blurred details. Specifically, gated images are often affected by special imaging conditions, frequently exhibiting low contrast, i.e., locally dark areas or locally overly bright areas due to laser aperture effects, resulting in the loss of detail information; simultaneously, in low-light environments, object outlines are blurred, and traditional convolution operations struggle to accurately capture clear edge features. To address these challenges, this invention provides targeted optimizations to the original self-calibrating convolution, proposing an extended spatial-channel convolution module, ESCConv, to better adapt to the feature extraction needs of gated images under complex lighting conditions.

[0059] like Figure 2As shown, the ESCConv module aims to integrate multi-scale feature extraction and high-frequency information enhancement. The input feature X (number of channels C) is divided into two parts, X1 and X2 (each with C / 2 channels), and the high-frequency component X3 is obtained through max pooling. X1 is processed by convolutional kernel K1 (with LeakReLU activation) to extract the backbone feature Y1. X2 is then processed sequentially by convolutional kernels K2, K3, and K4 (K4 using LeakReLU) with different receptive fields to form branch features Y2. The high-frequency feature X3 is compressed through a 1×1 convolution K5 and upsampled back to its original size before being fused with Y1 and Y2 in the channel dimension, outputting the final feature map Y. This structure effectively improves the ability to extract local edges and perceive global structures in low-light and high-noise environments.

[0060] The extended spatial-channel convolution module ESCConv comprehensively introduces multi-receptive-field convolution operations, high-frequency feature guidance, and multi-branch fusion strategies. Specifically, the input features are first extracted using max pooling to highlight edge details. Subsequently, the features are divided into two groups along the channel dimension. One group enters a multi-scale convolution branch to extract local and global information under different receptive fields, while the other group undergoes lightweight convolution to preserve the original feature distribution. Simultaneously, the high-frequency components are processed by channel compression and upsampling to restore the spatial scale before being fused with the features from the two main branches, ultimately yielding output features that combine detailed characterization with overall perception.

[0061] The ESCConv module effectively improves the detail recovery and structure preservation capabilities of feature maps under complex lighting conditions, making it particularly suitable for boundary extraction and contrast enhancement in low-light and noisy environments in gated images. To further explore the intrinsic dependencies between feature channels, the ResCAB module was added. Based on a three-stage operation of "compression-excitation-calibration," it dynamically assigns weights to the importance of features in each channel, strengthening key features and suppressing redundant interference. The specific process is as follows:

[0062] (1)

[0063] (2)

[0064] (3)

[0065] In formula (1), the spatial dimensions of the feature map X are H×W. The channel responses at a location are compressed into a one-dimensional channel vector through global average pooling to extract the overall feature distribution. Formula (2) introduces ReLU and Sigmoid nonlinear mapping, and the dependency relationship between features is modeled through a two-layer fully connected network (weight matrices W1, W2, compression ratio r). Finally, as shown in Formula (3), the activation coefficients are multiplied back into the original feature map element by element by channel to achieve targeted enhancement, highlight key channel features, and improve channel utilization efficiency and information expression ability. In order to fully integrate local and global features, a dual-skip connection path is designed in the feature extraction stage: one path directly transmits shallow low-frequency features to deep layers to maintain global structural information; the other path dynamically adjusts the importance weights of different channels based on the channel attention mechanism. The two are fused under the residual framework to effectively enhance the detail recovery and overall consistency in low-light and complex background environments. In addition, the encoder-decoder structure strengthens the information transmission between feature layers through skip connections to achieve collaborative modeling of high-level semantics and low-level details. At the end of feature extraction, a self-supervised module SSB is introduced to align the preliminary feature prediction with the ESCConv output, generating a pseudo-supervised signal. This improves discriminative feature extraction and feature quality in an unsupervised environment, laying the foundation for subsequent adaptive enhancement and fusion, and constructing an end-to-end optimization closed loop.

[0066] 1.2 Adaptive Feature Enhancement and Fusion Stage (ASFR)

[0067] In existing technologies, feature enhancement and fusion strategies suffer from the accumulation of noise interference during feature fusion and shortcomings in scale perception and region adaptive adjustment, making it difficult to effectively improve image quality in low-light and complex backgrounds. To address this issue, this invention provides an improved gated image feature enhancement and fusion module (ASFR), characterized by comprising an adaptive single-channel feature enhancement module (ASFE) and an adaptive spatial channel fusion module (ASCF), which work synergistically to optimize image feature enhancement and fusion. Specifically, the adaptive single-channel feature enhancement module (ASFE) enhances single-channel features, improving the expression of local detail information and effectively suppressing the negative impact of noise on image quality; the adaptive spatial channel fusion module (ASCF) dynamically adjusts the weights of space and channels to effectively fuse feature information at different scales, optimizing detail fidelity and overall consistency. Based on the FES-SCFM structure, this module further enhances and fuses features on top of the multi-scale high-quality features extracted in the EFES stage, improving the feature recovery capability and image quality of gated images under low-light and complex background conditions.

[0068] (1) Adaptive Single-channel Feature Enhancement (ASFE) module

[0069] The ASFE module combines local saliency perception with global structural constraints. Its key feature is its ability to explicitly enhance the extraction and expression of key features, suppress irrelevant interference information, and achieve ordered feature fusion, providing a highly robust feature foundation for subsequent illumination compensation and noise suppression stages. For example... Figure 3 As shown, the ASFE module adopts a multi-branch architecture, containing multiple convolutional branches at different scales. The outputs of each branch are concatenated along the channel dimension to generate a fused feature map. Each branch first uses a 1×1 convolution to adjust the number of channels and perform preliminary feature mapping on the input features. The fourth branch is set as a residual channel to ensure the complete preservation of the original feature information at the output, enhancing the ability to transmit fine-grained structural information. The other three branches use 1×3, 3×1, and 3×3 convolutional kernels, respectively, focusing on edge extraction, texture analysis, and local feature capture, thereby capturing rich image details at different scales.

[0070] After generating the fused feature map, Adaptive Contrast Enhancement (ACE) is introduced to dynamically weight the features. Through region-aware pooling, adaptive lightweight convolution, and non-linear activation operations, the dependencies between channels are effectively modeled, and the channel weights are dynamically adjusted according to the importance of local features. This design can improve the model's ability to focus on effective features in noisy environments while ensuring low computational cost, thereby significantly enhancing the feature representation and robustness of the gated image under complex low-light conditions.

[0071] (2) Adaptive Spatial-Channel Fusion (ASCF) module

[0072] To achieve efficient fusion of multi-scale features in gated images and enhance the dynamic synergy between local information representation and global structural integrity, an improved lightweight spatial channel fusion module is proposed. This module consists of a channel dimension selection mechanism and a spatial domain response adjustment mechanism, which respectively perform cross-channel information enhancement and region-sensitivity redistribution. The aim is to ensure that the output features simultaneously possess context-aware global perception and detail-level response capabilities. Its structure is as follows: Figure 4 As shown.

[0073] a. Channel Selection Unit (CA)

[0074] This unit adaptively models the input feature map along the channel direction to differentiate and highlight key semantics while suppressing invalid interference. The module input includes two feature branches originating from the ASFE module output: branch Y1' (representing local structure) and branch Y2' (representing overall layout). For the Y1' feature, Local Minimal Pooling (LMP) and Variable Receptive Field Pooling (ALP) strategies are introduced to extract boundary and texture responses at different scales, enhancing its fine-grained resolution. The Y2' branch focuses on structural stability and contextual modeling, employing Global Max Pooling (GMP) and Global Mean Pooling (GAP) for information compression and description. The outputs of these four pooling methods are concatenated along the channel direction to form a feature vector Z, which undergoes nonlinear mapping via a Multilayer Perceptron (MLP) and is normalized using a Sigmoid function to generate two sets of channel response weight coefficients, which are applied to Y1' and Y2' respectively. 1’ With the Y2' feature, channel-level adaptive reweighting is achieved. The mapping relationship is as follows:

[0075] (4)

[0076] (5)

[0077] in, For the Sigmoid function, and These are the channel attention weight mappings for local and global branches, respectively.

[0078] b. Spatial Response Unit SA

[0079] To further enhance the spatial responsiveness of the target region and achieve region-selective enhancement and suppression, this unit performs spatial dimension feature compensation based on channel-weighted results. Structurally, it also includes two processing paths, targeting the Y1' and Y2' feature maps respectively. The Y1' path first adjusts the channel dimension through 1×1 convolution to compress computational complexity, then introduces region-aware pooling (RAP) to generate a local response weight matrix K, improving the ability to identify abrupt feature regions. The Y2' path uses a joint modeling approach of global mean pooling and convolution to output a spatial-level weight vector V, enhancing macroscopic semantic consistency. Finally, K and V are multiplied to complete multi-scale information fusion. The result is then normalized by a combination of the Softmax and Sigmoid functions to form the final spatial attention distribution map, which is applied to the Y1' and Y2' features respectively, further refining the spatial feature selection.

[0080] To further enhance the salient location information in the feature maps and explore the interaction between local salient regions and global structural information, after completing the channel-dimensional feature recalibration, a combination of local detail awareness and global dependency modeling is used to effectively supplement the fine-grained spatial differences that channel attention cannot capture. The improved spatial attention (SA) consists of two branches. The first branch targets the local branch feature map Y1', using 1×1 convolution to adjust the channel dimension and combining it with region adaptive pooling (RAP) to generate a weight mapping matrix K, highlighting regional burst features and capturing key detail information. The second branch targets the global branch feature map Y2', using global average pooling (GAP) and convolution to jointly extract global spatial information, obtaining a weight vector V to enhance global dependencies and prevent information conflicts and repetitions. Subsequently, the local mapping matrix K and the global vector V are multiplied to fuse multi-scale spatial awareness information, forming a preliminary spatial weight distribution. The final spatial attention weights are then obtained through joint normalization using the Softmax and Sigmoid functions. The entire process can be represented as:

[0081] (6)

[0082] (7)

[0083] in, This is the Softmax function. After performing channel and spatial attention weighting, this module performs element-wise multiplication of the feature maps of the two branches with their corresponding channel and spatial weights, and then sums the results with weights to obtain the final fused feature representation. The fusion calculation formula is as follows:

[0084] (8)

[0085] here, This constitutes the weighted local feature Y1'. This constitutes the weighted global feature Y2''.

[0086] In this invention, after the ASFE outputs the weighted feature map Y', to achieve refined fusion of local and global features, the feature map is divided into two functionally complementary branches, generating Y1' and Y2' respectively. Specifically,

[0087] First, Y1' is transformed using two independent 1x1 convolution operations to reduce channel dimensional redundancy and highlight response characteristics at different scales.

[0088] One branch is processed by Region Adaptive Pooling (RAP) and subsequent convolutional layers to form Y1''. This branch focuses on fine-grained modeling of local regions and can capture local contrast, texture details and small object features.

[0089] Another branch extracts features from a larger receptive field directly through shallow convolutions to form Y2'', in order to maintain overall illumination consistency and global structural stability.

[0090] 1.3 Denoising Module ADNet

[0091] Gated imaging technology enables effective detection of distant targets by setting a laser pulse time window. However, in complex environments, gated images often suffer from insufficient contrast, loss of detail, and speckle noise due to uneven illumination, scattering interference, and noise, which greatly limits the improvement of image quality. To address these challenges, an Attention-guided Denoising Network (ADNet) is proposed for gated image denoising. ADNet effectively improves the denoising performance of gated images under harsh conditions by introducing a sparse block (SB), a feature enhancement block (FEB), an attention block (AB), and a reconstruction block (RB).

[0092] The SB module extracts noise features progressively through a 12-layer convolutional network, employing alternating combinations of dilated convolutions (Conv) and standard convolutions (Conv). The introduction of dilated convolutions (in layers 2, 5, 9, and 12) expands the receptive field, enabling the network to capture local and global noise distribution features in far-distance gated images. The remaining layers use standard convolutions to reduce computational overhead. The FEB module, based on a long-path structure, constructs a deep-to-shallow feature fusion path through a four-layer convolutional network. The first three layers use a Conv+BN+ReLU combination to enhance local detail representation; the fourth layer uses ordinary convolutions to integrate long-path features and perform feature fusion. The innovation of the FEB module lies in combining the Tanh activation function to map high-frequency noise information from the input image to a specific space, thereby improving sensitivity to far-distance noise signals.

[0093] To address the mixed distribution of noise and background details in gated images, the AB module generates attention weights through a 1×1 convolutional layer, thereby optimizing the extraction of noise features. Specifically, the AB module first compresses the output of the FEB module to generate a noise saliency map; then, it multiplies this map with the FEB output to enhance the representation of salient noise regions. This module is particularly suitable for noise suppression in backlit, highly reflective, and heavily scattered regions. The RB module combines the input noisy image with a noise map to restore a clear image. This module not only focuses on noise suppression but also preserves target edges and detail information, thus ensuring the overall visual quality of the gated image. When processing distant weak signal regions, the RB module effectively reduces speckle noise through repeated iterations of the noise map, while highlighting target contour features, thereby optimizing the sharpness and structural information of distant gated images.

[0094] 2. Experiment and Results Analysis

[0095] 2.1 Evaluation Indicators

[0096] Four evaluation metrics were introduced in the experiment: Peak Signal-to-Noise Ratio (PSNR), Natural Image Quality Assessment (NIQE), and Perceptual Similarity Index (LPIPS). PSNR quantitatively reflects the pixel-level error between the enhanced image and the reference image; a higher value indicates lower distortion. NIQE performs reference-free quality assessment based on a natural image statistical model; a lower value indicates a more natural image. LPIPS extracts perceptual features through a deep neural network to measure differences in images at the perceptual level; a smaller value indicates better perceptual quality.

[0097] 2.2 Experimental Environment and Parameter Settings

[0098] The model was trained and tested in an Ubuntu 18.04 operating system environment. The GPU was an Nvidia GeForce RTX 2080Ti with 11GB of VRAM. The model was based on the PyTorch framework, with Python version 3.9 and CUDA version 10.2. The batch size was set to 16, the training was conducted for 200 epochs, the learning rate was 0.0001, and the weight decay was 0.0001.

[0099] 2.3 Dataset

[0100] This invention combines a self-built dataset with publicly available datasets for training and evaluation to improve the model's adaptability and generalization ability in complex low-light scenes. The self-built dataset was acquired using a master-gated imaging system based on a 905nm laser band. This imaging system integrates a solid-state laser, a nanosecond-level gated camera, and a high-precision timing module. The laser emits a laser signal with a center wavelength of 905nm and a pulse width of 150µs. The camera employs an image-enhanced design, featuring a minimum 3ns gate width and 25W power consumption, achieving a synchronization accuracy of 35ps with the timing module, effectively capturing the instantaneous response information of the target scene. A total of 8082 images were acquired, covering various typical environments, and divided into training, validation, and test sets in an 8:1:1 ratio. To further verify the model's performance under natural low-light conditions, 312 images were randomly selected from the LOL publicly available dataset and also divided in the same ratio. Gated imaging samples are shown below. Figure 6 As shown.

[0101] 2.4 Comparative Experiment

[0102] This section provides a systematic comparative analysis of the improved gated image enhancement algorithm of this invention with several mainstream methods, covering traditional image enhancement methods (HE, MBLLEN) and deep learning-based methods (Retinex-Net, KinD++, SCI, EnlightenGAN, PairLIE, Zero-DCE++). To evaluate the enhancement effects of different methods on gated images, experiments were conducted in complex outdoor scenes. The visual comparison results are as follows: Figure 7 As shown.

[0103] In complex outdoor environments, existing enhancement methods often lead to overall brightness imbalance and even overexposure. Furthermore, due to limited model generalization ability, the enhanced image is prone to introducing noise over a wide area, further reducing visibility. In contrast, the improved enhancement algorithm in this paper effectively suppresses overexposure and noise amplification, optimizes brightness distribution, makes the enhanced image more visually natural, and preserves more high-frequency details, improving the resolution of distant targets.

[0104] Furthermore, to verify the robustness of the proposed method, low-light images were randomly selected from the LOL dataset for testing, and the results were visualized. Figure 8 On the LOL dataset, the proposed method can stably improve the global naturalness of images. The enhanced images significantly improve brightness balance and contrast while maintaining the original structural information, further verifying the generalization ability of the method.

[0105] Experimental results show that, compared with existing methods, the proposed method not only effectively suppresses artifacts, overexposure, and noise amplification, but also accurately preserves the image's tonal details during enhancement, improving overall naturalness. Notably, some contrasting methods, due to improper brightness adjustment or limited generalization ability, result in abnormal brightening or sudden noise increases in local areas. The proposed method, through refined feature extraction and enhancement strategies, effectively alleviates such distortions, consistently improving the visual quality of gated images under different data distributions, further optimizing image quality.

[0106] Several images were randomly selected from the gated image dataset and the LOL dataset for testing. To measure differences in pixel-level information, perceptual quality, and contrast, PSNR was used as the full-reference metric, while NIQE and LPIPS were used as no-reference metric. Higher PSNR values ​​indicated better image quality, while lower NIQE and LPIPS values ​​indicated better image quality. The performance of various enhancement methods, including HE, MBLLEN, Retinex-Net, KinD++, SCI, EnlightenGAN, PairLIE, and Zero-DCE++, was compared and evaluated.

[0107] Table 1 presents the quantitative evaluation results of different methods on the gated image dataset. Compared with the state-of-the-art method, the proposed method achieves a PSNR of 23.2388 dB, an improvement of 0.6579 dB. This indicates a significant reduction in distortion of the enhanced image. Furthermore, NIQE and LPIPS scores further validate the effectiveness of the method. The enhanced image significantly improves contrast while maintaining naturalness, reducing artifacts and over-enhancement issues, making the enhanced image closer to the real scene.

[0108] Table 1 Performance Evaluation of Gated Image Dataset Augmentation Algorithms

[0109]

[0110] To further validate the generalization ability of our proposed method, additional tests were conducted on the LOL dataset to evaluate its adaptability in different low-light scenarios. Table 2 shows the quantitative evaluation results of different methods on the LOL dataset. Experimental results show that our proposed method performs well in terms of PSNR, achieving a 0.7014 dB improvement compared to the best existing method. Furthermore, our proposed method also achieves good performance in NIQE and LPIPS metrics, effectively reducing image distortion and achieving significant enhancements in structure, contrast, and brightness consistency. In summary, the superior performance of our proposed method on multiple datasets validates its strong generalization ability, enabling it to robustly improve image quality under various low-light environments and provide more reliable input data for downstream tasks such as object detection.

[0111] Table 2 Performance Evaluation of LOL Data Augmentation Algorithms

[0112]

[0113] 2.5 Ablation Experiment

[0114] To verify the effectiveness of each module in the proposed method, we conducted a series of ablation experiments on the gated image dataset and the LOL dataset to quantitatively analyze the contribution of each module to the overall enhancement framework. The quantitative results are shown in Tables 3 and 4.

[0115] Table 3 Performance evaluation of gated dataset ablation experiments

[0116]

[0117] We progressively introduced different improved modules (EFES, ASFE, ASCF, ADNet) into the base augmentation network to explore their impact on the final augmentation performance. Specifically, in the feature extraction stage, we replaced the original FES with the EFES module to further improve feature extraction capabilities. EFES captures additional local information through the MaxPooling mechanism and dynamically adjusts the weights of different channels using a channel attention mechanism to enhance key features. Furthermore, the introduction of the LeakReLU activation function improves the ability to adjust the feature distribution, enabling the network to dynamically adjust the negative slope based on the feature distribution of different layers during training. This avoids excessive suppression of gradient information in dark areas and improves the enhancement capability of low-light details. Experimental results show that the EFES module improves the PSNR by 0.5491 dB on the gated dataset and by 0.4562 dB on the LOL dataset.

[0118] Building upon this foundation, an adaptive feature enhancement and fusion (ASFR) module (ASFE+ASCF) is further introduced to enhance feature fusion capabilities. This module combines local and global features, employing channel and spatial optimization mechanisms to effectively reduce information loss while enhancing feature representation. Experiments show that the ASFR module improves PSNR by 0.5661 dB on the gated dataset and by 0.7805 dB on the LOL dataset.

[0119] Table 4 Performance Evaluation of Ablation Experiments on the LOL Dataset

[0120]

[0121] Finally, a denoising module, ADNet, is added to further enhance the robustness of the augmentation network. This module utilizes a noise-aware gating mechanism to adaptively determine the noise distribution of the input features, thereby specifically suppressing non-uniform noise and improving the enhancement quality of the gated image. Experimental results show that this module improves the PSNR by 0.2975 dB on the gated dataset and by 0.6515 dB on the LOL dataset. In summary, the ablation experiments demonstrate that each improved module of the proposed method positively contributes to the enhancement performance, ultimately forming a gated image enhancement framework with stronger robustness and generalization ability.

[0122] 3. Conclusion

[0123] To address the limitations of gated images in complex environments such as high noise interference, uneven illumination, and loss of high-frequency details, this paper proposes an end-to-end enhancement method to systematically improve image quality and visual perception capabilities. The ESCConv feature extraction module expands the receptive field using a multi-branch structure and introduces max-pooling to enhance edge and high-frequency information, effectively compensating for the shortcomings of traditional methods in detail preservation and edge sharpening. The ASFR adaptive feature enhancement and fusion module enhances the expressive power of texture details and improves the adaptability of contrast adjustment by dynamically balancing global and local features. The IAS illumination adjustment module implements adaptive illumination compensation, significantly improving the problem of information loss in dark areas caused by uneven illumination distribution. The ADNet noise module, based on a modular design strategy, not only effectively removes speckle noise but also significantly improves the contrast and detail clarity of the gated image, thus providing more accurate input data for subsequent object detection and recognition tasks.

[0124] References

[0125] [1] Wang Shuyu, Ai Lei, Tao Shengxiang, et al. Research on portable long-range laser gating imaging system [J]. Journal of Ordnance Equipment Engineering, 2018, 39(09):166-170.

[0126] [2] Wang Xinwei, Sun Liang, Zhang Yue, et al. Research progress on laser distance-gated three-dimensional imaging technology (invited) [J]. Infrared and Laser Engineering, 2024, 53(04):31-49.

[0127] [3]X. Guo, Y. Li and H. Ling, LIME: Low-Light Image Enhancement via Illumination Map Estimation, IEEE Transactions on Image Processing, 26(2):982-993, 2017.2

[0128] [4]M. Li, J. Liu, W. Yang, X. Sun and Z. Guo, Structure-RevealingLow-Light Image Enhancement Via Robust Retinex Model. IEEE Transactions onImage Processing,27(6):2828-2841,2018.2

[0129] [5]Li C , Guo J , Porikli F ,et al.LightenNet: A Convolutional NeuralNetwork for weakly illuminated image enhancement[J].Pattern recognitionletters, 2018, 104(MAR.1):15-22.

[0130] [6]Shijie H ,Xu H ,Yanrong G , et al.Low-Light Image Enhancement withSemi-Decoupled Decomposition[J].IEEE Transactions on Multimedia,2020,1-1.

[0131] [7]Wang W L ,Liu S Z ,Siu C W , et al.Lightening Network for Low-light Image Enhancement[J].IEEE Transactions on Image Processing,2020,1-1.

[0132] [8]Shakibania H ,Raoufi S ,Khotanlou H .CDAN: Convolutional denseattention-guided network for low-light image enhancement[J].Digital SignalProcessing,2025,156(PA):104802-104802.

Claims

1. A long-range gating image enhancement method based on an improved JCRNet, characterized by: First, the joint correction and thinning network JCRNet is improved to obtain the gating image enhancement algorithm; The gated image is then enhanced using a gated image enhancement algorithm. The processing steps include: 1) In the feature extraction stage of JCRNet, the residual block is replaced by the enhanced self-calibrating convolution ESCConv module; at the same time, LeakReLU activation is used to improve the nonlinear mapping capability; The feature extraction steps for a given input image X with C input channels include: 1.1) Divide X into X1 and X2 with C / 2 channels; at the same time, X is max-pooled to form a high-frequency component X3; 1.2) X1 is transformed by convolution kernel K1 to obtain the main branch feature Y1; X2 is fed into the ESCConv module and processed by three convolution kernels K2, K3 and K4 with different receptive fields in sequence to finally obtain the output Y2; X3 is first processed by 1×1 convolution K5 for channel compression, and then restored to the original size by Upsample to obtain Y3; K1 and K4 both use the LeakReLU activation function; 1.3) Y1, Y2 and Y3 are concatenated to obtain the final output feature map Y; 2) In the joint refinement stage of JCRNet, the gated image feature enhancement and fusion module ASFR is used to process Y; ASFR consists of the adaptive single-channel feature enhancement module ASFE and the adaptive spatial channel fusion module ASCF. The processing steps for Y include: 2.1) In ASFE, firstly, Y is processed by multiple convolutional branches of different scales, and the processing results are concatenated along the channel dimension to form a fused feature map; then, the adaptive channel enhancement ACE mechanism is used to weight the feature map to obtain Y'; the weighted feature map Y' is divided into two functionally complementary branches to generate local features Y1' and global features Y2' respectively; 2.2) ASCF includes channel attention (CA) branch and spatial attention (SA) branch, which respectively process the local feature Y1' and global feature Y2' output by ASFE. The steps include: 2.2.1) In the CA branch, for Y1', local max pooling (LMP) and adaptive local pooling (ALP) are used to extract local features respectively; for Y2', global max pooling (GMP) and global average pooling (GAP) are used to capture overall features. The four pooling results are concatenated into a fused feature description vector Z, which is then subjected to a nonlinear transformation via a multilayer perceptron (MLP) and normalized using a sigmoid function to obtain the channel attention weight mappings corresponding to Y1' and Y2', respectively. and ; 2.2.2) In the SA branch, Y1' first extracts preliminary features through 1×1 convolution, and then applies region adaptive pooling (RAP) to generate a spatial weight matrix K; Y2' extracts global spatial features through global average pooling (GAP) combined with the convolution module to form a weight vector V. Subsequently, K and V are multiplied by matrix, and then normalized using a combination of Softmax and Sigmoid to obtain the final spatial attention weights. and ; and Corresponding to Y1' and Y2' respectively; 2.2.3) Perform element-wise multiplication of the two branches Y1' and Y2' with their corresponding channel and spatial weights respectively, and then sum the results in a weighted manner to obtain the final fusion feature. : 。 2. The long-range gating image enhancement method based on improved JCRNet according to claim 1, characterized in that: The result of the input image X processed by the IAS lighting adjustment module of JCRNet is added to the result processed in step 2), and then further processed by the attention-guided denoising network ADNet. ADNet consists of sparse blocks (SB), feature enhancement blocks (FEB), attention blocks (AB), and reconstruction blocks (RB). For input noisy images, ADNet's processing steps include: 3.1) The SB module extracts noise features step by step through a 12-layer convolutional network, where layers 2, 5, 9 and 12 are dilated convolutions (Conv), and the remaining layers use standard convolutions (Conv). 3.2) The FEB module is based on a long path structure and constructs a deep and shallow feature fusion path through a four-layer convolutional network; the first three layers use a combination of Conv+BN+ReLU, and the fourth layer uses a regular convolutional Conv; the FEB module uses the Tanh activation function. 3.3) The AB module first performs weighted compression on the output of the FEB module through a 1×1 convolutional layer to generate a noise saliency map; then the noise saliency map is multiplied with the FEB output to obtain a noise mapping map; 3.4) The RB module combines the input noisy image with the noise map to restore a clear image.

3. The long-range gating image enhancement method based on the improved JCRNet according to claim 1, characterized in that: The processing of X2 in the ESCConv module includes: 1.2.1) The results of X2 being processed sequentially by AvgPooling, K3, and Upsample are added to X2; 1.2.2) X2 is processed by K2; 1.2.3) After activation processing, the result of 1.2.1) is multiplied by the result of 1.2.2), and then processed by K4 to obtain Y2.

4. The long-range gating image enhancement method based on improved JCRNet according to claim 1, characterized in that: In step 2.1), the ASFE has four convolutional branches: the first branch has a 1×1 convolution; the second branch has 1×1, 3×1, 1×3 and 3×3 convolutions in sequence; the third branch has 1×1, 1×3, 3×1 and 3×3 convolutions in sequence; and the fourth branch has 1×1 and 3×3 convolutions in sequence, which serve as residual paths. The results of Y processed by the second, third and fourth branches are concatenated and then added to the result of the first branch to obtain the fused feature map.

5. The long-range gating image enhancement method based on the improved JCRNet as described in claim 1, Its characteristic is that in the feature extraction stage, the ResCAB module is used to further process the output of the ESCConv module. After the processing result is processed by the encoder-decoder module, it is then processed by the self-supervised module SSB to align the preliminary features with the features output by ESCConv and generate a pseudo-supervised signal. The encoder and decoder in the encoder-decoder module are connected by a jumper.

Images