A visual auxiliary active RIS beam prediction method for abnormal light environment

Abstract

This invention relates to a vision-assisted active RIS beam prediction method for abnormal lighting environments, belonging to the field of wireless communication. This method, designed for wireless communication systems, uses a camera mounted on the active RIS side to acquire image data of the communication environment under abnormal lighting conditions. A wireless beam prediction system model is constructed, and a deep neural network model and a training strategy are employed to learn from the acquired image data and predict the optimal communication beam. This invention improves the performance of the active RIS prediction model in abnormal lighting environments, accelerates the convergence speed of the network model, enhances the accuracy of beam prediction, and strengthens the robustness and reliability of the active RIS wireless communication system in abnormal lighting conditions.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication and relates to a visual-assisted active RIS beam prediction method for abnormal lighting environments. Background Technology

[0002] Wireless communication, with its advantages of high bandwidth and low latency, supports efficient data transmission and is widely used in next-generation communication systems such as 5G and 6G. However, the transmission distance of wireless signals is limited and easily obstructed by obstacles, often leading to challenges in the reliability and stability of communication links. Reconfigurable smart surfaces, as an innovative wireless communication technology, have attracted widespread attention due to their ability to intelligently control the wireless propagation environment in a low-power and low-cost manner. However, in large-scale RIS deployments, channel estimation involves extremely high parameter dimensions, significantly increasing pilot overhead and computational complexity, severely restricting the system's real-time response capability. To address this challenge, RIS communication schemes based on sensing data assistance are gradually gaining importance. These schemes leverage artificial intelligence to analyze the changing patterns of environmental sensing data, enabling efficient prediction of the optimal communication beam in the active RIS beamcodebook.

[0003] Video and image data contain communication-related information such as user location and are widely used in wireless beam prediction tasks. However, in abnormal lighting environments with insufficient or unstable light, image acquisition quality deteriorates significantly, leading to a substantial reduction in the accuracy of vision-based beam prediction. Taking industrial manufacturing scenarios as an example, complex or limited lighting conditions significantly increase the difficulty of vision-assisted wireless prediction, thereby causing data transmission instability. Therefore, improving the accuracy of beam prediction under abnormal lighting conditions has become a critical issue that urgently needs to be addressed. Summary of the Invention

[0004] Technical Problem: In view of this, the purpose of this invention is to provide a visual-assisted active RIS beam prediction method for abnormal lighting environments. It addresses the problem of visual-assisted wireless beam prediction in abnormal lighting scenarios by employing a course training method to ensure high-accuracy beam prediction in abnormal lighting communication environments.

[0005] Technical Solution: To achieve the above objectives, the present invention provides the following technical solution:

[0006] A vision-assisted active RIS beam prediction method for abnormal lighting environments is proposed. Addressing the increased beam prediction difficulty caused by reduced wireless communication information in image data under abnormal lighting conditions, a course-based training-based beam prediction method is presented. Difficult samples (i.e., data collected under abnormal lighting conditions) are introduced into the training set in stages according to their difficulty, guiding the prediction model to learn in an order from easy to difficult, thereby effectively improving the beam prediction accuracy of the model in abnormal lighting environments.

[0007] The method specifically includes the following steps:

[0008] S1: Construct an active RIS wireless communication system. Use a neural network model to extract features from image data under abnormal lighting conditions such as low light, overexposure, uneven light, sudden light change, and stray light. Optimize the beamforming vector and select the optimal communication beam from the active RIS beamcodebook to maximize the received signal power and reduce the complexity of beam prediction.

[0009] S2: Collect image data of the active RIS wireless communication scenario and the corresponding optimal communication beam data in the active RIS beam codebook and form a dataset. Preprocess the image data collected by the active RIS camera. Based on the collected dataset, construct an active RIS beam prediction deep neural network model and define the model loss function and model optimizer to train the model.

[0010] The preprocessing of image data acquired by the active RIS camera specifically involves: dividing the image dataset acquired by the active RIS camera into different categories of abnormal light image subsets and normal light image subsets; adding the different categories of abnormal light image datasets to the normal light image training set in stages based on the course training method; using the normal light image dataset as the original training set in the first stage of training; and uniformly adding abnormal light images to the training set in subsequent stages; and finally predicting the optimal beam in the entire dataset.

[0011] S3: Input the entire dataset into the trained active RIS beam prediction deep neural network model to predict the optimal active RIS beam.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] S11: Construct an active RIS wireless communication system, comprising a mobile user equipment equipped with an omnidirectional antenna to receive downlink signals from a communication base station; an active RIS smart metasurface equipped with a camera for capturing real-time environmental images; and a communication base station for receiving and parsing environmental images from the active RIS camera and distributing the active RIS beamcodebook. The active RIS includes a dynamic smart reflector composed of M reflective elements, and beamforming vectors are used to... The signal is reflected back to the mobile user equipment, where Represents the beamforming vector in the codebook, codebook , This refers to the number of beam vectors in the codebook. Active RIS wireless communication systems employ orthogonal frequency division multiplexing (OFDM) technology, through... The active RIS uses multiple subcarriers to transmit signals. In the reflection link between the active RIS and the user, the communication channel for each subcarrier is represented as follows: ,in ;

[0014] S12: Define the beam prediction optimization problem and predict the optimal beamforming vector using a neural network model. To maximize the received signal power on a given subcarrier, the optimization objective formula is:

[0015]

[0016] Beam prediction targets can be defined from the beam codebook Select the optimal communication beam to connect and communicate with the user; if This represents an image of a communication scene captured by an active RIS camera. This represents the optimal beam predicted by the neural network. If we denote the real number space, then the beam prediction function can be expressed as:

[0017]

[0018] in, The parameters represent the optimization parameters of the network model. Since there is a one-to-one correspondence between the beam vector and its index in the codebook, the prediction difficulty can be further reduced by directly predicting the beam index. Therefore, the original formula can be expressed as:

[0019] , ;

[0020] in, This represents the index of the predicted optimal beam vector.

[0021] Furthermore, in step S2, an active RIS beam prediction deep neural network model is constructed and trained, specifically including the following steps:

[0022] S21: The active RIS uses a camera to capture real-time environmental images in RGB format, containing multiple feature information of the communication environment. The active RIS transmits the captured image information back to the main serving communication base station, which then uses a preset active RIS beamcodebook. Select the optimal beamforming vector The beamcodebook contains multiple beamforming vectors. To cover the entire communication scenario; the collected dataset includes both normal and abnormal light data sets;

[0023] S22: Construct an active RIS beam prediction deep neural network model. Use the pre-trained neural network model ResNet50 to perform the beam prediction task. Input the acquired image into the ResNet network model. The ResNet network model processes the image data and extracts deep features of the image through multi-layer convolution and pooling operations.

[0024] S23: After image feature extraction is completed, the active RIS beamforming prediction deep neural network model uses fully connected layers to map the features to the beamforming vector space, and calculates the probability distribution of each beamforming vector using the Softmax function. The training of the active RIS beamforming prediction deep neural network model adopts the cross-entropy loss function, and the optimization objective is to maximize the probability of correctly predicting the beamforming vector. The loss function is expressed as:

[0025]

[0026] in, For loss function, The number of training samples, For the model in a given image In the case of predicting the beam vector index To further improve the generalization ability of the active RIS beam prediction deep neural network model, the Adam optimizer is used to dynamically adjust the parameters of the active RIS beam prediction deep neural network model to ensure that the model converges quickly during training.

[0027] S24: Select the beamforming vector with the highest probability as the output result.

[0028] Furthermore, in step S2, the preprocessing of the acquired image data specifically includes the following steps:

[0029] S201: Divide the training dataset into easy samples and hard samples, where easy samples include images taken under normal lighting conditions, and hard samples include images taken under abnormal lighting conditions. D represents the segmented normal light image dataset. unnormal This represents the segmented abnormal lighting image dataset, which includes K subsets of abnormal lighting data such as low light, overexposure, uneven lighting, abrupt changes in lighting, and stray light. This represents the low-light dataset within the abnormal lighting dataset; the sample distributions of the normal lighting image dataset and the abnormal lighting image dataset are respectively... P unnormal Correspondingly, the distribution of low-light dataset samples in the abnormal lighting dataset is represented as follows: ;

[0030] S202: The course training method is adopted, in which the training process starts with image data under normal lighting conditions and gradually introduces images taken under abnormal lighting conditions to improve the beam prediction accuracy of the neural network model under abnormal lighting conditions. By adding abnormal lighting data to the training set in stages, the network model's ability to learn from complex abnormal lighting samples is increased.

[0031] Specifically, the training process is divided into K categories of abnormal light, each category containing In each stage, each category completes a full training session on the abnormal lighting dataset for that category. Each stage also adds a more challenging abnormal lighting dataset of the same category to the training set. Taking the training of the low-light abnormal lighting dataset as an example, let... Indicates the first The probability of random sampling added to the training set from abnormal light samples at each stage; using Indicates the first Training sets for each stage, Indicates the first The distribution of training set samples at each stage, then and It can be represented as:

[0032]

[0033]

[0034] in, Indicates a random sample dataset Total In the new dataset that is formed, the training set in the final stage will contain all the normal and abnormal light datasets.

[0035] set up and This indicates that in the first stage of training, all normal light images were used as the initial training set to train the model, and in subsequent stages, abnormal light samples were uniformly added to the training set. A new training set is formed by adding elements to the training set; finally, the optimal beam is predicted on the entire dataset.

[0036] The beneficial effects of this invention are as follows:

[0037] (1) This invention takes into account the problem of reduced accuracy of active RIS smart metasurface wireless beam prediction under abnormal lighting conditions, optimizes the prediction under abnormal lighting conditions, and ensures the stability and reliability of 5G wireless communication.

[0038] (2) The present invention adopts the course training method to improve beam prediction performance. By utilizing the phased learning method of course training, the model can learn more general features from simple samples, which effectively improves the accuracy of model prediction and enables industrial 5G communication to maintain stable and reliable implementation. Attached Figure Description

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0040] Figure 1 This is a model of the vision-assisted active RIS communication system provided by the present invention;

[0041] Figure 2 This is a schematic diagram of deep learning training based on course training according to the present invention;

[0042] Figure 3 This is a flowchart of the visual-assisted active RIS beam prediction method based on course training according to the present invention. Detailed Implementation

[0043] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0044] A visually assisted active RIS beam prediction method for abnormal lighting environments specifically includes the following steps:

[0045] S1: Construct an active RIS wireless communication system. Use a neural network model to extract features from image data under abnormal lighting conditions such as low light, overexposure, uneven light, sudden light change, and stray light. Optimize the beamforming vector and select the optimal communication beam from the active RIS beamcodebook to maximize the received signal power and reduce the complexity of beam prediction.

[0046] S2: Collect image data of the active RIS wireless communication scenario and the corresponding optimal communication beam data in the active RIS beam codebook and form a dataset. Preprocess the image data collected by the active RIS camera. Based on the collected dataset, construct an active RIS beam prediction deep neural network model and define the model loss function and model optimizer to train the model.

[0047] The preprocessing of image data acquired by the active RIS camera specifically involves: dividing the image dataset acquired by the active RIS camera into different categories of abnormal light image subsets and normal light image subsets; adding the different categories of abnormal light image datasets to the normal light image training set in stages based on the course training method; using the normal light image dataset as the original training set in the first stage of training; and uniformly adding abnormal light images to the training set in subsequent stages; and finally predicting the optimal beam in the entire dataset.

[0048] S3: Input the entire dataset into the trained active RIS beam prediction deep neural network model to predict the optimal active RIS beam.

[0049] Step S1 specifically includes the following steps:

[0050] S11: Construct an active RIS wireless communication system, comprising a mobile user equipment equipped with an omnidirectional antenna to receive downlink signals from a communication base station; an active RIS smart metasurface equipped with a camera for capturing real-time environmental images; and a communication base station for receiving and parsing environmental images from the active RIS camera and distributing the active RIS beamcodebook. The active RIS includes a dynamic smart reflector composed of M reflective elements, and beamforming vectors are used to... The signal is reflected back to the mobile user equipment, where Represents the beamforming vector in the codebook, codebook , This refers to the number of beam vectors in the codebook. Active RIS wireless communication systems employ orthogonal frequency division multiplexing (OFDM) technology, through... The active RIS uses multiple subcarriers to transmit signals. In the reflection link between the active RIS and the user, the communication channel for each subcarrier is represented as follows: ,in ;

[0051] S12: Define the beam prediction optimization problem and predict the optimal beamforming vector using a neural network model. To maximize the received signal power on a given subcarrier, the optimization objective formula is:

[0052]

[0053] Beam prediction targets can be defined from the beam codebook Select the optimal communication beam to connect and communicate with the user; if This represents an image of a communication scene captured by an active RIS camera. Let the optimal beam predicted by the neural network be represented, then the beam prediction function can be expressed as:

[0054]

[0055] in, The parameters represent the optimization parameters of the network model. Since there is a one-to-one correspondence between the beam vector and its index in the codebook, the prediction difficulty can be further reduced by directly predicting the beam index. Therefore, the original formula can be expressed as:

[0056] , ;

[0057] in, This represents the index of the predicted optimal beam vector.

[0058] In step S2, the active RIS beam prediction deep neural network model is constructed and trained, specifically including the following steps:

[0059] S21: The active RIS uses a camera to capture real-time environmental images in RGB format, containing multiple feature information of the communication environment. The active RIS transmits the captured image information back to the main serving communication base station, which then uses a preset active RIS beamcodebook. Select the optimal beamforming vector The beamcodebook contains multiple beamforming vectors. To cover the entire communication scenario; the collected dataset includes both normal and abnormal light data sets;

[0060] S22: Construct an active RIS beam prediction deep neural network model. Use the pre-trained neural network model ResNet50 to perform the beam prediction task. Input the acquired image into the ResNet network model. The ResNet network model processes the image data and extracts deep features of the image through multi-layer convolution and pooling operations.

[0061] S23: After image feature extraction is completed, the active RIS beamforming prediction deep neural network model uses fully connected layers to map the features to the beamforming vector space, and calculates the probability distribution of each beamforming vector using the Softmax function. The training of the active RIS beamforming prediction deep neural network model adopts the cross-entropy loss function, and the optimization objective is to maximize the probability of correctly predicting the beamforming vector. The loss function is expressed as:

[0062]

[0063] in, For loss function, The number of training samples, For the model in a given image In the case of predicting the beam vector index To further improve the generalization ability of the active RIS beam prediction deep neural network model, the Adam optimizer is used to dynamically adjust the parameters of the active RIS beam prediction deep neural network model to ensure that the model converges quickly during training.

[0064] S24: Select the beamforming vector with the highest probability as the output result.

[0065] In step S2, the acquired image data is preprocessed, specifically including the following steps:

[0066] S201: Divide the training dataset into easy samples and hard samples, where easy samples include images taken under normal lighting conditions, and hard samples include images taken under abnormal lighting conditions. D represents the segmented normal light image dataset. unnormal This represents the segmented abnormal lighting image dataset, which includes K subsets of abnormal lighting data such as low light, overexposure, uneven lighting, abrupt changes in lighting, and stray light. This represents the low-light dataset within the abnormal lighting dataset; the sample distributions of the normal lighting image dataset and the abnormal lighting image dataset are respectively... P unnormal Correspondingly, the distribution of low-light dataset samples in the abnormal lighting dataset is represented as follows: ;

[0067] S202: The course training method is adopted, in which the training process starts with image data under normal lighting conditions and gradually introduces images taken under abnormal lighting conditions to improve the beam prediction accuracy of the neural network model under abnormal lighting conditions. By adding abnormal lighting data to the training set in stages, the network model's ability to learn from complex abnormal lighting samples is increased.

[0068] Specifically, the training process is divided into K categories of abnormal light, each category containing In each stage, each category completes a full training session on the abnormal lighting dataset for that category. Each stage also adds a more challenging abnormal lighting dataset of the same category to the training set. Taking the training of the low-light abnormal lighting dataset as an example, let... Indicates the first The probability of random sampling added to the training set from abnormal light samples at each stage; using Indicates the first Training sets for each stage, Indicates the first The distribution of training set samples at each stage, then and It can be represented as:

[0069]

[0070]

[0071] in, Indicates a random sample dataset Total In the new dataset that is formed, the training set in the final stage will contain all the normal and abnormal light datasets.

[0072] set up and This indicates that in the first stage of training, all normal light images were used as the initial training set to train the model, and in subsequent stages, abnormal light samples were uniformly added to the training set. A new training set is formed by adding elements to the training set; finally, the optimal beam is predicted on the entire dataset.

[0073] See Figures 1-3 This invention provides a vision-assisted active RIS beam prediction method for abnormal lighting environments. For vision-assisted active RIS communication systems, a camera installed at the active RIS end captures real-time abnormal lighting images of the communication environment to predict which communication beam from the codebook the active RIS should select to communicate with the user. A ResNet50 network model is used for beam prediction, and a course-based training method is employed to accelerate training speed and improve prediction accuracy, thus inventing a deep learning-based beam prediction scheme.

[0074] Figure 1This is a model diagram of a vision-assisted active RIS communication system. It includes a fixed base station, a RIS reflector, a mobile user, and other moving scattering objects and obstacles. The active RIS is equipped with a camera and a dynamic intelligent reflector, with M reflector elements. The active RIS captures and collects real-time images of the communication environment, forming a raw dataset. In this system, the active RIS uses a predefined beam steering codebook. Connect and communicate with users.

[0075] Figure 2 This is a diagram of a deep neural network prediction framework based on course training. The entire training process is divided into K categories of abnormal light rays, each category containing... Each stage contains [number] phases, and each phase includes [number] phases. Number of training epochs. Each category completes one full training iteration on the abnormal lighting dataset for that category. In each stage, normal light images are initially used as the training set, and in subsequent stages, abnormal light datasets of the same category are gradually and randomly added to the training set. For the beam prediction task, a deep residual network, ResNet50, is used for prediction. In the final classification output layer, the predicted output category is fine-tuned to the number of beam codebooks. When image data is input into the model, the convolutional layers and residual blocks in ResNet50 extract depth features from the image, and then pass them through fully connected layers to generate the predicted output.

[0076] Figure 3 This is a flowchart of the visual-assisted active RIS beam prediction method for abnormal lighting environments according to the present invention. Taking low light as an example, for a certain type of abnormal lighting data, the method for predicting the optimal beam of the active RIS based on course training and deep neural networks specifically includes the following steps:

[0077] V1~V4: Obtain relevant datasets for beam prediction. Divide the collected datasets into two subsets: abnormal light image datasets for low light category and normal light image datasets. Use these subsets to initialize model hyperparameters, the number of training stages, and the number of training rounds per stage for course training.

[0078] V4~V10: Build the ResNet50 network model for beam training and define the loss function and optimizer in the network model. Use normal light images as the original training set for the first stage of training. In the subsequent training stages, gradually and evenly add low-light category abnormal light image datasets for course training.

[0079] V10~V15: Network training and prediction process. During the training process, the course is implemented in stages. The input image trains the entire network, the loss function is calculated and the network parameters are updated by backpropagation. After the training termination condition is reached, the optimal predicted beam is finally output.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A vision-aided active RIS beam prediction method for non-line-of-sight environments, characterized in that, Specifically, the following steps are included: S1: Construct an active RIS wireless communication system, extract features from image data under abnormal lighting conditions using a neural network model, optimize beamforming vectors, and select the optimal communication beam from the active RIS beamcodebook to maximize received signal power; S2: Collect image data of active RIS wireless communication scenarios and the optimal communication beam data in the active RIS beam codebook and form a dataset. Preprocess the image data collected by the active RIS camera. Based on the collected dataset, construct an active RIS beam prediction deep neural network model and define the model loss function and model optimizer to train the model. The preprocessing of image data acquired by the active RIS camera specifically involves: dividing the image dataset acquired by the active RIS camera into different categories of abnormal light image subsets and normal light image subsets; adding the different categories of abnormal light image datasets to the normal light image training set in stages based on the course training method; using the normal light image dataset as the original training set in the first stage of training; and uniformly adding abnormal light images to the training set in subsequent stages; and finally predicting the optimal beam in the entire dataset. S3: Input the entire dataset into the trained active RIS beam prediction deep neural network model to predict the optimal active RIS beam.

2. The vision-aided active RIS beam prediction method for non-line-of-sight (NLOS) light environment according to claim 1, wherein, Step S1 specifically includes the following steps: S11: Construct an active RIS wireless communication system, comprising a mobile user equipment equipped with an omnidirectional antenna to receive downlink signals from a communication base station; an active RIS smart metasurface equipped with an active RIS camera for capturing real-time environmental images; and a communication base station for receiving and parsing the environmental images from the active RIS camera and distributing the active RIS beamcodebook. The active RIS smart metasurface includes a dynamic smart reflector composed of M reflective elements, and is beamformed by a beamforming vector. The signal is reflected back to the mobile user equipment, where Represents the beamforming vector in the codebook, codebook , It is the number of beam vectors in the codebook; the active RIS wireless communication system uses orthogonal frequency division multiplexing technology, through The active RIS uses multiple subcarriers to transmit signals. In the reflection link between the active RIS and the user, the communication channel for each subcarrier is represented as follows: ,in ; S12: Define the beam prediction optimization problem and predict the optimal beamforming vector using a neural network model. To maximize the received signal power on a given subcarrier, the optimization objective formula is: ； The beam prediction target is defined from the active RIS beam codebook Select the optimal communication beam to connect and communicate with the user; if This represents an image of a communication scene captured by an active RIS camera. If the optimal beam predicted by the neural network is represented, then the beam prediction function is expressed as: ； in, The parameters represent the optimization parameters of the network model. Since there is a one-to-one correspondence between the beam vectors and their indices in the codebook, the prediction difficulty is reduced by directly predicting the beam indices. Therefore, the original formula can be expressed as: ， ； in, This represents the index of the predicted optimal beam vector.

3. The visual-assisted active RIS beam prediction method for abnormal lighting environments according to claim 2, characterized in that, In step S2, the active RIS beam prediction deep neural network model is constructed and trained, specifically including the following steps: S21: The active RIS uses a camera to capture real-time environmental images in RGB format, containing multiple feature information of the communication environment. The active RIS transmits the captured image information back to the main serving communication base station, which then uses a preset active RIS beamcodebook. Select the optimal beamforming vector The beamcodebook contains multiple beamforming vectors. To cover the entire communication scenario; the collected dataset includes both normal and abnormal light data sets; S22: Construct an active RIS beam prediction deep neural network model. Use the pre-trained neural network model ResNet50 to perform the beam prediction task. Input the acquired image into the ResNet network model. The ResNet network model processes the image data and extracts deep features of the image through multi-layer convolution and pooling operations. S23: After image feature extraction is completed, the active RIS beamforming prediction deep neural network model uses fully connected layers to map the features to the beamforming vector space, and calculates the probability distribution of each beamforming vector using the Softmax function. The training of the active RIS beamforming prediction deep neural network model adopts the cross-entropy loss function, and the optimization objective is to maximize the probability of correctly predicting the beamforming vector. The loss function is expressed as: ； in, For loss function, The number of training samples, For the model in a given image In the case of predicting the beam vector index To improve the generalization ability of the active RIS beam prediction deep neural network model, the Adam optimizer is used to dynamically adjust the parameters of the active RIS beam prediction deep neural network model, ensuring that the active RIS beam prediction deep neural network model converges quickly during training. S24: Select the beamforming vector with the highest probability as the output result.

4. The visual-assisted active RIS beam prediction method for abnormal lighting environments according to claim 2, characterized in that, In step S2, the acquired image data is preprocessed, specifically including the following steps: S201: Divide the training dataset into easy samples and hard samples, where easy samples include images taken under normal lighting conditions, and hard samples include images taken under abnormal lighting conditions. D represents the segmented normal light image dataset. unnormal This represents the segmented abnormal lighting image dataset, which includes K subsets of abnormal lighting data: low light, overexposure, uneven lighting, abrupt changes in lighting, and stray light. This represents the low-light dataset within the abnormal lighting dataset; the sample distributions of the normal lighting image dataset and the abnormal lighting image dataset are respectively... and P unnormal The distribution of low-light dataset samples in the abnormal lighting dataset is represented as follows: ; S202: The course training method is adopted, in which the training process starts with image data under normal lighting conditions and gradually introduces images taken under abnormal lighting conditions to improve the beam prediction accuracy of the neural network model under abnormal lighting conditions. By adding abnormal lighting data to the training set in stages, the network model's ability to learn from complex abnormal lighting samples is increased. The training process is divided into K categories of abnormal light, each category containing... In each of the three phases, a complete training exercise is performed on the abnormal lighting dataset for that category. At each phase, a more challenging abnormal lighting dataset for that category is added to the training set. Indicates the first The probability of random sampling added to the training set from abnormal light samples at each stage; using Indicates the first Training sets for each stage, Indicates the first The distribution of training set samples at each stage, then and Represented as: ； ； in, Indicates a random sample dataset Total In the new dataset that is formed, the training set in the final stage will contain all the normal and abnormal light datasets. set up and This indicates that in the first stage of training, all normal light images were used as the initial training set to train the model, and in subsequent stages, abnormal light samples were uniformly added to the training set. A new training set is formed by adding elements to the training set; finally, the optimal beam is predicted on the entire dataset.

Images