Double-subnet low-illumination enhancement method for night monitoring scene of expressway

Abstract

The invention relates to a double-subnet low-illumination image enhancement method in a high-speed scene. The method comprises the following steps: S1, collecting night low-illumination image data of an expressway through a high-speed camera; s2, a multi-stage cascade framework is constructed, the multi-stage cascade framework comprises a difference calibration sub-network and an illumination channel self-weighting sub-network, and an unsupervised loss function is used to train the multi-stage cascade framework; and S3, in a test stage, removing the difference calibration sub-network in the multi-stage cascade framework, processing an input low-illumination image only by using the illumination channel self-weighting sub-network, and outputting an enhanced image. According to the invention, high-quality dark light enhancement of the night traffic monitoring image of the expressway can be realized, and the overall brightness and definition of the image can be obviously improved under the condition of low illuminance. The CSWM is introduced into the illumination channel self-weighted sub-network, so that the overexposure problem caused by vehicle headlamps, pavement marking lines and reflective objects is effectively relieved.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a dual-subnet low-light enhancement method for highway nighttime monitoring scenarios. Background Technology

[0002] With the continuous growth of highway traffic flow and the widespread adoption of intelligent traffic management systems, nighttime video surveillance plays an increasingly important role in traffic safety management and emergency response. However, under low-light conditions, highway surveillance images generally suffer from two prominent problems: firstly, insufficient overall brightness leads to a loss of detail in dark areas, accompanied by amplified sensor noise, severely affecting the identification of key features such as vehicle outlines, license plate numbers, and road markings; secondly, overexposed areas caused by strong light sources such as vehicle headlights and streetlights further obscure license plate and traffic sign information. These problems directly reduce the accuracy and reliability of nighttime traffic monitoring in vehicle detection, traffic incident analysis, and accident location.

[0003] While existing low-light image enhancement methods have made some progress, they still fall short of meeting the practical needs of highway nighttime monitoring. Traditional methods based on Retinex theory rely on manual parameter adjustments, making them ill-suited to high-speed dynamic traffic flow and complex lighting environments, easily resulting in overexposure in bright areas and color distortion in dark areas. Supervised learning methods, although capable of fitting complex lighting maps, heavily depend on paired low-light-normal-light training samples. However, in real traffic scenarios, it is almost impossible to acquire such paired data on a large scale, and synthetic data struggles to cover extreme environments such as tunnel entrances and rainy nights, leading to insufficient model generalization ability. Unsupervised methods alleviate the dependence on paired data to some extent, but generally suffer from low computational efficiency and a lack of modeling for traffic scene-specific noise, making them difficult to deploy in real-time in monitoring systems with high traffic volume and demanding emergency response requirements.

[0004] In summary, existing technologies generally suffer from insufficient noise suppression in dark areas, inadequate handling of overexposure in strong light, and a lack of real-time performance in low-light environments. These issues directly lead to the loss of critical information such as vehicle details, license plates, and traffic signs, thus affecting the efficiency and accuracy of traffic enforcement, accident tracing, and emergency management. Therefore, there is an urgent need to propose an image enhancement method that can be trained under unsupervised conditions, effectively balance noise suppression and overexposure control in low-light environments, and possess lightweight and real-time processing capabilities to meet the stringent application requirements of nighttime traffic monitoring on highways. Summary of the Invention

[0005] In view of this, the present invention aims to propose a dual-subnet low-light enhancement method for highway nighttime monitoring scenarios to solve the problems in the prior art.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention proposes a dual-subnet low-light enhancement method for highway nighttime monitoring scenarios, the method comprising the following steps: S1, which uses a high-speed camera to collect low-light image data of highways at night; S2, Construct a multi-stage cascaded framework, which includes a differential calibration subnetwork and an illumination channel self-weighted subnetwork, and train the multi-stage cascaded framework using an unsupervised loss function; S3, During the testing phase, the difference calibration subnetwork in the multi-stage cascaded framework is removed, and only the illumination channel self-weighting subnetwork is used to process the input low-light image, outputting an enhanced image.

[0007] Furthermore, in step S1, the collected image data includes images of typical highway sections, which include at least one of straight sections, curves, entrances and exits, toll stations, and tunnel entrances / exits. The data collection time covers the low-light periods after sunset, late-night peak hours, and before dawn; the image data contains dynamic traffic elements, including at least one of the following: high-speed vehicles, license plates from multiple angles, direct headlight illumination, street lighting, tunnel light and dark boundaries, road markings, traffic signs, and sudden emergency events.

[0008] Furthermore, step S1 also includes data enhancement of the acquired image data, including spatial transformation enhancement and illumination interference enhancement; the spatial transformation enhancement includes rotation and mirroring operations; the illumination interference enhancement includes adjusting brightness and contrast to simulate at least one of dark light, vehicle headlight glare, streetlight shadows, and mixed light interference.

[0009] Furthermore, the structure of the difference calibration subnetwork includes: The input layer receives the difference representation between the original image and the output of the previous stage; A combination of convolutional layers, batch normalization layers, and ReLU activation layers; Three residual blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation layer; Wavelet domain module (WDM) is used to separate noise from details in the frequency domain; The output layer contains convolutional layers and a sigmoid activation function, and outputs a calibration map.

[0010] Furthermore, the wavelet domain module (WDM) captures different frequency domain features through multi-scale decomposition, including three branches: the upper branch retains key information through standard convolution, and the lower two branches extract multi-scale features and restore the dimensions through inverse wavelet transform before fusing them with the upper branch; Its output calculation formula is: in, To output new or eigenvalues, For inverse discrete wavelet transform, X is the input image of the module. (⋅) represents a subnetwork that performs adaptive processing on the features of the j-th layer and the s-th subband. These are learnable weights assigned to the j-th layer and the s-th subband, used to highlight important frequency components. It is a learnable residual connection coefficient used to control the degree to which the original input information is preserved.

[0011] Furthermore, the structure of the self-weighted sub-network of the illumination channel includes: The input layer receives low-light images or difference calibration outputs. Convolutional layers and ReLU activation layers; Three residual blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation layer; Convolutional layers and Sigmoid activation layers; Channel Self-Weighted Module (CSWM) is used to suppress overexposed channels and correct colors.

[0012] Furthermore, the Channel Self-Weighted Module (CSWM) obtains global channel information through global average pooling, learns the dependencies between channels through a fully connected layer, and generates weights to weight the input channels. Its output calculation formula is: Where A is the input of CSWM, B is the output of CSWM, and λ is a learnable fusion coefficient. This represents element-wise multiplication, implementing channel-weighted multiplication. It is the Sigmoid activation function. This is a global average pooling operation that compresses the input A into a scalar feature vector of the channel dimension for subsequent processing. This is a channel attention mechanism that uses a fully connected network to weight GAP features, thereby enhancing the response of important channels. This is a non-local attention mechanism (NL module) that captures global dependencies and improves feature representation through cross-locational interactions. This is the residual convolution branch, which extracts local features and preserves the original information through convolution operations, preventing gradient vanishing.

[0013] Furthermore, the unsupervised loss function includes fidelity loss and smoothness loss, and the total loss function is: in, It's a loss of fidelity. It is the smoothness loss, and α and β are two positive balance parameters.

[0014] Furthermore, the formula for calculating the fidelity loss is as follows: Where T is the total number of stages. This represents the illumination output at stage n, where y is the original low-light input. This is the output of the difference calibration module; The formula for calculating the smoothness loss is: Where M is the total number of pixels, and i represents the i-th pixel. This represents the adjacent pixels of i within its 5×5 window. It is a key spatial adaptive weight that determines the penalty strength for the difference in illumination between adjacent pixels.

[0015] Furthermore, in step S3, the processing formula for the testing phase is as follows: Where y is the raw low-light image input from the roadside camera, L is the self-weighted subnet of the illumination channel, and Z is the network processing output.

[0016] Compared with the prior art, the present invention has the following advantages: This invention enables high-quality low-light enhancement of highway nighttime traffic monitoring images, significantly improving overall image brightness and clarity under low-light conditions. By introducing CSWM into the self-weighted subnetwork of the illumination channel, overexposure problems caused by vehicle headlights, road markings, and reflective objects are effectively alleviated, and color distortion is improved. Simultaneously, WDM is integrated into the difference calibration subnetwork, and high-frequency noise and low-frequency background are separated through multi-branch processing in the frequency domain. This effectively suppresses graininess in dark areas and sensor noise while preserving key details such as vehicle outlines and road markings. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 Flowchart for training a multi-stage cascaded framework; Figure 2 This is a diagram of the differential calibration subnetwork structure of the present invention; Figure 3 This is a structural diagram of the WDM of the present invention; Figure 4 This is a diagram of the self-weighted subnetwork structure of the illumination channel of the present invention; Figure 5This is a structural diagram of the CSWM of the present invention; Figure 6 This is a structural diagram of the testing phase of the present invention; Figure 7 This is a test scenario diagram for the present invention; Figure 8 This is a scene image after low-light processing according to the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0019] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "back," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0020] Furthermore, in the description of this invention, unless otherwise explicitly defined, the terms "installation," "connection," "linking," and "connector" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention in light of the specific circumstances.

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] Overall, this invention proposes a dual-subnet low-light enhancement method for highway nighttime monitoring scenarios, the method comprising the following steps: S1, which uses a high-speed camera to collect low-light image data of highways at night; S2, Construct a multi-stage cascaded framework, which includes a differential calibration subnetwork and an illumination channel self-weighted subnetwork, and train the multi-stage cascaded framework using an unsupervised loss function; S3, during the testing phase, removes the differential calibration subnetwork in the multi-stage cascaded framework and uses only the illumination channel self-weighting subnetwork to process the input low-light image, outputting an enhanced image.

[0023] It should be noted that in step S1, the collected image data includes images of typical sections of the highway, and typical sections include at least one of straight sections, curves, entrances and exits, toll stations, and tunnel entrances / exits; The data collection time covers the low-light periods after sunset, late night peak hours, and before dawn; the image data contains dynamic traffic elements, including at least one of the following: high-speed vehicles, license plates at multiple angles, direct headlights, street lighting, tunnel light and shadow boundaries, road markings, traffic signs, and sudden emergency events.

[0024] Step S1 also includes data augmentation of the acquired image data, including spatial transformation enhancement and illumination interference enhancement; spatial transformation enhancement includes rotation and mirroring operations; illumination interference enhancement includes adjusting brightness and contrast to simulate at least one of dark light, vehicle headlight glare, streetlight shadows and mixed light interference.

[0025] In this embodiment, to construct a dedicated dataset suitable for low-light image enhancement tasks on highways at night, image data is first collected using surveillance cameras deployed at key traffic points on typical highway sections, such as straight sections, curves, entrances and exits, toll stations, and tunnel entrances / exits. The data collection time covers the periods with the most pronounced low-light characteristics, including after sunset, late-night rush hour, and the low-light transition period before dawn. To ensure data diversity and representativeness, special attention is paid to segments containing dynamic traffic elements during data collection, such as different speed states of high-speed vehicles, various license plates from multiple angles and under different lighting conditions (direct headlight illumination, street lighting, tunnel light / dark transition), the visibility of road markings and traffic signs, and sudden emergency events (such as breakdowns, accident scenes). Subsequently, keyframes containing rich traffic information are extracted from the video stream as raw RGB images. Based on these, traffic scene adaptive data enhancement is performed, including spatial transformation enhancement and illumination interference enhancement. Spatial transformation enhancement simulates differences in camera installation angles and vehicle travel direction changes through operations such as rotation and mirroring. Illumination interference enhancement adjusts brightness and contrast to simulate varying degrees of low light, headlight glare, streetlight shadows, and mixed light interference. Finally, the processed image data is divided into training, test, and validation sets in a ratio of 7:2:1, forming a dedicated low-light nighttime dataset for highways covering various road types, traffic conditions, and lighting conditions, providing a sample foundation for subsequent model training and performance evaluation.

[0026] To suppress noise interference in enhanced images, this invention proposes a difference calibration subnetwork in the framework; to reduce local overexposure and color deviation problems in enhanced images, this invention proposes a self-weighted illumination channel subnetwork in the framework.

[0027] During the training phase, the first stage only includes a self-weighted illumination subnetwork. Using low-light images from roadside cameras as input, it learns the illumination residuals to perform initial brightness and color correction. Subsequent stages dynamically generate calibration maps through a difference calibration subnetwork, adding them pixel-by-pixel to the initial input image. This preserves the original structural information while suppressing noise and correcting details. A second illumination optimization is then performed through the self-weighted illumination subnetwork L. This cascaded design progressively drives the outputs of each stage to converge to a consistent result, ensuring the stability and consistency of the enhanced image in terms of brightness, color, and structural detail.

[0028] Among them, such as Figure 2 As shown, the structure of the difference calibration subnetwork includes: an input layer that receives the difference representation between the original image and the output of the previous stage; a combination of convolutional layers, batch normalization layers, and ReLU activation layers; three residual blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation layer; a wavelet domain module (WDM) for separating noise and details in the frequency domain; and an output layer containing convolutional layers and a sigmoid activation function to output the calibration map.

[0029] like Figure 3 As shown, the wavelet domain module (WDM) captures different frequency domain features through multi-scale decomposition, including three branches: the upper branch retains key information through standard convolution, and the lower two branches extract multi-scale features and restore the dimensions through inverse wavelet transform before fusing them with the upper branch; The core of this module utilizes multi-scale decomposition to capture features in different frequency domains. It comprises three branches: the upper branch preserves key information through standard convolution, while the lower two branches extract multi-scale features, restore the dimensionality through inverse wavelet transform, and then fuse them with the upper branch to optimize output smoothness. The output calculation formula is as follows: in, To output new or eigenvalues, For inverse discrete wavelet transform, X is the input image of the module. (⋅) represents a subnetwork that performs adaptive processing on the features of the j-th layer and the s-th subband. These are learnable weights assigned to the j-th layer and the s-th subband, used to highlight important frequency components. It is a learnable residual connection coefficient used to control the degree to which the original input information is preserved.

[0030] The structure of the illumination channel self-weighted subnetwork includes: an input layer that receives low-light images or difference calibration outputs; convolutional layers and ReLU activation layers; three residual blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation layer; convolutional layers and sigmoid activation layers; and a channel self-weighted module (CSWM) for suppressing overexposed channels and correcting colors.

[0031] like Figure 5 As shown, the Channel Self-Weighted Module (CSWM) obtains global channel information through global average pooling, learns the dependencies between channels through a fully connected layer, and generates weights to weight the input channels. Based on this module, global channel information is obtained through global average pooling, and the fully connected layer learns the dependencies between channels to generate weights to suppress overexposed channels and correct colors. The output calculation formula is as follows: Where A is the input of CSWM, B is the output of CSWM, and λ is a learnable fusion coefficient. This represents element-wise multiplication, implementing channel-weighted multiplication. It is the Sigmoid activation function. This is a global average pooling operation that compresses the input A into a scalar feature vector of the channel dimension for subsequent processing. This is a channel attention mechanism that uses a fully connected network to weight GAP features, thereby enhancing the response of important channels. This is a non-local attention mechanism (NL module) that captures global dependencies and improves feature representation through cross-locational interactions. This is the residual convolution branch, which extracts local features and preserves the original information through convolution operations, preventing gradient vanishing.

[0032] This module automatically learns the importance weights of feature channels and highlights key channel features through weighted summaries, thereby mitigating overexposure and color distortion issues.

[0033] The unsupervised loss function includes fidelity loss and smoothness loss, and the total loss function is: in, It's a loss of fidelity. It is the smoothness loss, and α and β are two positive balance parameters.

[0034] To elaborate further, the formula for calculating fidelity loss is: Where T is the total number of stages. This represents the illumination output at stage n, where y is the original low-light input. This is the output of the difference calibration module; using the redefined input. To constrain the output lighting Instead of relying on handcrafted ground-based data or simple low-light input, we employ an L1-norm smoothing term with spatial variability, given that illumination smoothing is a general consensus for low-light enhancement tasks. The formula for calculating smoothness loss is: Where M is the total number of pixels, and i represents the i-th pixel. This represents the adjacent pixels of i within its 5×5 window. It is a key spatial adaptive weight that determines the penalty strength for the difference in illumination between adjacent pixels.

[0035] The model is trained entirely independently of paired training data, enabling it to adapt to diverse scenarios. Simultaneously, the two unique subnet designs allow the framework to achieve high-quality low-light image enhancement while remaining lightweight.

[0036] like Figure 6 As shown, in step S3, the processing formula for the testing phase is: Where y is the raw low-light image input from the roadside camera, L is the self-weighted subnet of the illumination channel, and Z is the network processing output.

[0037] in addition, Figure 7 This paper presents the original input image of a typical nighttime monitoring scenario on a highway. This image presents significant challenges under low-light conditions, particularly with severe overexposure in the streetlight area, leading to the loss of crucial details. Simultaneously, the overall image exhibits heavy noise and graininess, affecting subsequent traffic information recognition and analysis. During the testing phase, this method addresses these types of original images by processing them through a self-weighted sub-network based on the illumination channel. This effectively suppresses overexposed areas, corrects color distortion, and reduces noise amplification in dark areas. Figure 8 The image shows an enhanced image processed by the present invention, with significantly improved visual quality, a marked reduction in overexposure, more natural color reproduction, clearer details of road surface texture, and easier identification of road condition information.

[0038] This invention enables high-quality low-light enhancement of highway nighttime traffic monitoring images, significantly improving overall image brightness and clarity under low-light conditions. By introducing CSWM into the self-weighted subnetwork of the illumination channel, overexposure caused by vehicle headlights, road markings, and reflective objects is effectively mitigated, and color distortion is improved. Simultaneously, WDM is integrated into the difference calibration subnetwork, achieving separation of high-frequency noise and low-frequency background through multi-branch processing in the frequency domain. This effectively suppresses graininess in dark areas and sensor noise while preserving key details such as vehicle outlines and road markings.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dual-subnet low-light enhancement method for highway nighttime monitoring scenarios, characterized in that: The method includes the following steps: S1, which uses a high-speed camera to collect low-light image data of highways at night; S2, Construct a multi-stage cascaded framework, which includes a differential calibration subnetwork and an illumination channel self-weighted subnetwork, and train the multi-stage cascaded framework using an unsupervised loss function; S3, During the testing phase, the difference calibration subnetwork in the multi-stage cascaded framework is removed, and only the illumination channel self-weighting subnetwork is used to process the input low-light image, outputting an enhanced image.

2. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 1, characterized in that: In step S1, the collected image data includes images of typical highway sections, which include at least one of straight sections, curves, entrances and exits, toll stations, and tunnel entrances / exits. The data collection time covers the low-light periods after sunset, late-night peak hours, and before dawn; the image data contains dynamic traffic elements, including at least one of the following: high-speed vehicles, license plates from multiple angles, direct headlight illumination, street lighting, tunnel light and dark boundaries, road markings, traffic signs, and sudden emergency events.

3. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 1, characterized in that: Step S1 further includes data enhancement of the acquired image data, including spatial transformation enhancement and illumination interference enhancement; the spatial transformation enhancement includes rotation and mirroring operations; the illumination interference enhancement includes adjusting brightness and contrast to simulate at least one of dark light, vehicle headlight glare, street light shadows and mixed light interference.

4. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 1, characterized in that: The structure of the differential calibration subnetwork includes: The input layer receives the difference representation between the original image and the output of the previous stage; A combination of convolutional layers, batch normalization layers, and ReLU activation layers; Three residual blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation layer; Wavelet domain module (WDM) is used to separate noise from details in the frequency domain; The output layer contains convolutional layers and a sigmoid activation function, and outputs a calibration map.

5. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 4, characterized in that: The wavelet domain module (WDM) captures different frequency domain features through multi-scale decomposition, including three branches: the upper branch retains key information through standard convolution, and the lower two branches extract multi-scale features and restore the dimensions through inverse wavelet transform before fusing them with the upper branch; Its output calculation formula is: in, To output new or eigenvalues, For inverse discrete wavelet transform, X is the input image of the module. (⋅) represents a subnetwork that performs adaptive processing on the features of the j-th layer and the s-th subband. These are learnable weights assigned to the j-th layer and the s-th subband, used to highlight important frequency components. It is a learnable residual connection coefficient used to control the degree to which the original input information is preserved.

6. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 1, characterized in that: The structure of the self-weighted subnetwork of the illumination channel includes: The input layer receives low-light images or difference calibration outputs. Convolutional layers and ReLU activation layers; Three residual blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation layer; Convolutional layers and Sigmoid activation layers; Channel Self-Weighted Module (CSWM) is used to suppress overexposed channels and correct colors.

7. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 6, characterized in that: The Channel Self-Weighted Module (CSWM) obtains global channel information through global average pooling, learns the dependencies between channels through a fully connected layer, and generates weights to weight the input channels. Its output calculation formula is: Where A is the input of CSWM, B is the output of CSWM, and λ is a learnable fusion coefficient. This represents element-wise multiplication, implementing channel-weighted multiplication. It is the Sigmoid activation function. This is a global average pooling operation that compresses the input A into a scalar feature vector of the channel dimension for subsequent processing. This is a channel attention mechanism that uses a fully connected network to weight GAP features, thereby enhancing the response of important channels. This is a non-local attention mechanism (NL module) that captures global dependencies and improves feature representation through cross-locational interactions. This is the residual convolution branch, which extracts local features and preserves the original information through convolution operations, preventing gradient vanishing.

8. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 1, characterized in that: The unsupervised loss function includes fidelity loss and smoothness loss, and the total loss function is: in, It's a loss of fidelity. It is the smoothness loss, and α and β are two positive balance parameters.

9. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 8, characterized in that: The formula for calculating the fidelity loss is: Where T is the total number of stages. This represents the illumination output at stage n, where y is the original low-light input. This is the output of the difference calibration module; The formula for calculating the smoothness loss is: Where M is the total number of pixels, and i represents the i-th pixel. This represents the adjacent pixels of i within its 5×5 window. It is a key spatial adaptive weight that determines the penalty strength for the difference in illumination between adjacent pixels.

10. The dual-subnet low-light enhancement method for highway nighttime monitoring scenarios according to claim 1, characterized in that: In step S3, the processing formula for the testing phase is as follows: Where y is the raw low-light image input from the roadside camera, L is the self-weighted subnet of the illumination channel, and Z is the network processing output.

Images