A light chasing light illumination angle and intensity calculation method based on a CNN deep learning algorithm

Abstract

The application discloses a light pursuit light illumination angle and intensity calculation method based on a CNN deep learning algorithm, and specifically comprises the following steps: collecting image information; pre-processing the image information; constructing a multi-branch deep network based on a CNN deep learning algorithm; and predicting an optimal light pursuit angle and a light intensity value through a full connection layer. Through the iteration and optimization of a reinforcement learning module control strategy, high-precision tracking is realized.

Description

Technical Field

[0001] This invention relates to the fields of intelligent control and optical detection technology, and in particular to a method for calculating the illumination angle and intensity of light tracking based on a CNN deep learning algorithm. Background Technology

[0002] Existing technology discloses a method, apparatus, storage medium, and electronic device for real-time ray tracing. The method includes: inputting an image of a target location in a scene to be rendered into pre-trained first, second, and third target adversarial neural network models to obtain stereo texture data of the target location. The stereo texture data includes: a normal stereo texture, a world coordinate stereo texture, and an albedo stereo texture; and calculating real-time lighting information from the stereo texture data. Some embodiments of this application can reduce computational load and achieve real-time ray tracing.

[0003] However, existing methods cannot calculate the optimal illumination angle and intensity while tracking dynamic light sources in real time. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed to provide a method for calculating light intensity based on a CNN deep learning algorithm to overcome or at least partially solve the above problems.

[0005] According to one aspect of the present invention, a method for calculating the angle and intensity of light tracing based on a CNN deep learning algorithm is provided, wherein the angle and intensity calculation method specifically includes:

[0006] Acquire image information;

[0007] The image information is preprocessed;

[0008] Constructing a multi-branch deep network based on the CNN deep learning algorithm;

[0009] Predict the optimal tracking angle and light intensity value using a fully connected layer.

[0010] Optionally, the image information includes: RGB image, near-infrared band, and time-series light intensity signal.

[0011] Optionally, the preprocessing of the image information specifically includes:

[0012] The RGB image is preprocessed;

[0013] The time-series light intensity signal is preprocessed.

[0014] Optionally, the preprocessing of the RGB image specifically includes:

[0015] Color space conversion: Converts BGR format images to LAB color space; the luminance channel L is calculated using a non-linear function to distinguish between highlight and low-brightness areas; the chrominance channels A and B are calculated using the difference to determine different color components.

[0016] Data augmentation: Add Gaussian noise to enhance model robustness;

[0017] Standardization: Normalization is performed based on the mean and standard deviation of the ImageNet dataset.

[0018] Optionally, the noise intensity of the Gaussian noise is 0.05.

[0019] Optionally, the preprocessing of the time-series optical intensity signal specifically includes:

[0020] Savitzky-Golay filters are used to smooth the light intensity timing signal and eliminate high-frequency noise.

[0021] The filter has a window length of 15 and a polynomial order of 3.

[0022] Optionally, the construction of a multi-branch deep network based on the CNN deep learning algorithm specifically includes:

[0023] Spatial Feature Branch: The input is a sequence of 5 consecutive stacked images, and spatiotemporal features are extracted through a 3D convolutional layer;

[0024] Dilated convolution is introduced to expand the receptive field, and convolution kernel parameters are dynamically generated, including those based on global average pooling feature vectors and generated by a multilayer perceptron.

[0025] Light intensity temporal branch: The input is the light intensity temporal signal with a 10-second window. Local features are extracted through a 1D convolutional layer, and long-term dependencies are modeled through a gated recurrent unit. Temporal features are then weighted and fused through an attention mechanism.

[0026] Adaptive weighted fusion layer: After concatenating spatial and temporal features, dynamic weights are generated through the Sigmoid function, and the two types of features are weighted and fused to output 160-dimensional fused features.

[0027] This invention provides a method for calculating the illumination angle and intensity of light tracking based on a CNN deep learning algorithm. The method specifically includes: acquiring image information; preprocessing the image information; constructing a multi-branch deep network based on the CNN deep learning algorithm; and predicting the optimal light tracking angle and intensity value through fully connected layers. High-precision tracking is achieved by iteratively optimizing the control strategy through a reinforcement learning module.

[0028] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart illustrating a method for calculating the illumination angle and intensity of light tracking based on a CNN deep learning algorithm, provided in an embodiment of the present invention;

[0031] Figure 2 This is a diagram of the multi-branch network structure of the CNN deep learning algorithm provided in an embodiment of the present invention.

[0032] Figure 3 This is a flowchart of the dynamic convolution kernel generation module provided in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the training and deployment process provided in an embodiment of the present invention. Detailed Implementation

[0034] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0035] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0036] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0037] like Figure 1 As shown, a method for calculating the angle and intensity of light tracing based on a CNN deep learning algorithm is described. The specific calculation method includes:

[0038] Acquire image information;

[0039] The image information is preprocessed;

[0040] Constructing a multi-branch deep network based on the CNN deep learning algorithm;

[0041] Predict the optimal tracking angle and light intensity value using a fully connected layer.

[0042] It integrates a multispectral camera and a light intensity sensor to acquire RGB images (resolution 1600×1200, 30 frames / second), near-infrared band (900-1700nm), and time-series light intensity signals (sampling rate 10Hz, range 0-88k Lux).

[0043] Image preprocessing includes:

[0044] a. Color Space Conversion: Converts BGR format images to LAB color space. The luminance channel (L) is calculated using a non-linear function to distinguish between highlight and low-brightness areas; the chroma channel (A / B) calculates different color components using the difference.

[0045] b. Data augmentation: Add Gaussian noise (noise intensity 0.05) to enhance model robustness.

[0046] c. Standardization: Normalization is performed based on the mean and standard deviation of the ImageNet dataset.

[0047] Timing signal preprocessing

[0048] A Savitzky-Golay filter (window length 15, polynomial order 3) is used to smooth the light intensity timing signal and eliminate high-frequency noise.

[0049] (3) CNN Model: The specific structure of the CNN model is as follows Figure 2 As shown.

[0050] Constructing a multi-branch deep network includes the following core modules:

[0051] a. Spatial Feature Branch (3D CNN): The spatial feature branch takes a sequence of 5 consecutive stacked images as input and extracts spatiotemporal features through 3D convolutional layers. Dilated convolution (dilation rate 2) is introduced to expand the receptive field, and convolutional kernel parameters are dynamically generated (based on global average pooling feature vectors, generated by a multilayer perceptron).

[0052] The flowchart of the dynamic convolution kernel generation module is as follows: Figure 3 As shown.

[0053] b. Light Intensity Temporal Branch (1D CNN+GRU): The input is a 10-second window of light intensity temporal signal. Local features are extracted through 1D convolutional layers, long-term dependencies are modeled through gated recurrent units (GRU), and finally, temporal features are weighted and fused through an attention mechanism.

[0054] c. Adaptive weight fusion layer: After concatenating spatial and temporal features, dynamic weights are generated through the Sigmoid function, and the two types of features are weighted and fused to output 160-dimensional fused features.

[0055] The output predicts the optimal tracking angle (azimuth 0°-360°) and light intensity value through a fully connected layer, and generates control commands.

[0056] Basic lighting images and light intensity temporal signals are input into the convolutional layer;

[0057] The input data is normalized, noise filtered, and multimodal data fused.

[0058] The selected spatial features are convolved in three dimensions, and the selected temporal features are convolved in one dimension. A dynamic two-branch feature weight mode is set.

[0059] Output the optimal tracking angle, predicted light intensity value, and control commands.

[0060] The training and deployment process is as follows: Figure 4 As shown.

[0061] Core algorithms include:

[0062] (1) Network Structure

[0063] Hybrid dilated convolution is used to enhance feature extraction capabilities, and adaptive parameterized ReLU (PReLU) is combined to improve nonlinear expression capabilities.

[0064] (2) Loss Function

[0065] The light intensity prediction (mean squared error loss) and angle prediction (cosine similarity loss) are jointly optimized with initial weights of 0.7 and 0.3, respectively, and the weights are adjusted by a dynamic optimization algorithm (DARTS).

[0066] (3) Training strategy

[0067] a. Transfer learning: Based on the SolarDB pre-trained model of the public illumination dataset.

[0068] b. Online incremental learning: During the deployment phase, model parameters are updated through the Elastic Weight Consolidation (EWC) strategy to avoid catastrophic forgetting.

[0069] Data acquisition uses a moving axis equipped with a multispectral camera to collect illumination data under various weather conditions, and enhances data diversity through random occlusion and brightness perturbation.

[0070] The hardware uses an embedded GPU, NVIDIA Jetson Nano, with an inference speed of ≤50 milliseconds; the software accelerates inference through TensorRT and supports ONNX model conversion.

[0071] Comparative experiments show that the light tracking accuracy of this invention reaches 98.2% under complex lighting conditions, which is 23% higher than that of the traditional PID control method; energy consumption is reduced by 40%, making it suitable for scenarios such as solar tracking systems and intelligent agricultural light control.

[0072] Beneficial effects:

[0073] (1) Multi-scale feature pyramid

[0074] By fusing spatial features at different resolutions through downsampling and upsampling, and combining them with temporal feature expansion, the ability to represent multimodal data is enhanced.

[0075] (2) Dynamic convolution kernel generation module

[0076] The convolution kernel parameters are adaptively adjusted according to the input illumination distribution to improve the flexibility of feature extraction.

[0077] (3) Light intensity-angle joint optimization

[0078] High-precision tracking is achieved by iteratively optimizing the control strategy through reinforcement learning modules.

[0079] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for calculating the illumination angle and intensity of light tracking based on a CNN deep learning algorithm, characterized in that, The method for calculating the angle and intensity specifically includes: Acquire image information; The image information is preprocessed; Multi-branch deep networks are constructed based on the CNN deep learning algorithm, including: Spatial Feature Branch: The input is a sequence of 5 consecutive stacked images, and spatiotemporal features are extracted through a 3D convolutional layer; Dilated convolution is introduced to expand the receptive field, and convolution kernel parameters are dynamically generated, including those based on global average pooling feature vectors and generated by a multilayer perceptron. Light intensity temporal branch: The input is the light intensity temporal signal with a 10-second window. Local features are extracted through a 1D convolutional layer, and long-term dependencies are modeled through a gated recurrent unit. Temporal features are then weighted and fused through an attention mechanism. Adaptive weight fusion layer: After concatenating spatial and temporal features, dynamic weights are generated through the Sigmoid function, and the two types of features are weighted and fused to output 160-dimensional fused features; Predict the optimal tracking angle and light intensity value using a fully connected layer.

2. The method for calculating the illumination angle and intensity of light tracking based on the CNN deep learning algorithm according to claim 1, characterized in that, The image information includes: RGB image, near-infrared band and time-series light intensity signal.

3. The method for calculating the illumination angle and intensity of light tracking based on the CNN deep learning algorithm according to claim 2, characterized in that, The preprocessing of the image information specifically includes: The RGB image is preprocessed; The time-series light intensity signal is preprocessed.

4. The method for calculating the illumination angle and intensity of light tracking based on the CNN deep learning algorithm according to claim 3, characterized in that, The preprocessing of the RGB image specifically includes: Color space conversion: Converts BGR format images to LAB color space; the luminance channel L is calculated using a non-linear function to distinguish between highlight and low-brightness areas; the chrominance channels A and B are calculated using the difference to determine different color components. Data augmentation: Add Gaussian noise to enhance model robustness; Standardization: Normalization is performed based on the mean and standard deviation of the ImageNet dataset.

5. The method for calculating the illumination angle and intensity of light tracking based on the CNN deep learning algorithm according to claim 4, characterized in that, The noise intensity of the Gaussian noise is 0.

05.

6. The method for calculating the illumination angle and intensity of light tracking based on the CNN deep learning algorithm according to claim 3, characterized in that, The preprocessing of the time-series light intensity signal specifically includes: Savitzky-Golay filters are used to smooth the light intensity timing signal and eliminate high-frequency noise. The filter has a window length of 15 and a polynomial order of 3.

Images