A multi-scene lamp decoration automatic control system based on deep learning
By combining U-Net neural networks and Q-reinforcement learning, we have achieved accurate classification and adaptive adjustment of a multi-scene lighting automatic control system. This solves the problems of inaccurate scene recognition and low user preference matching in traditional systems, thus improving the user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG LEISHISHAN TECHNOLOGY GROUP CO LTD
- Filing Date
- 2026-03-25
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional multi-scene automatic lighting control systems lack effective mining of spatial structure information in deep images, have coarse semantic segmentation granularity, high classification confusion rate, and do not fully integrate information complexity and spatial structure features, resulting in a mismatch between lighting control parameters and actual scene requirements, and a poor user experience. At the same time, the fixed parameter mapping mode does not consider users' personalized operating habits, has weak adaptive adjustment capabilities, and is difficult to balance objective scene requirements with users' personalized preferences.
A U-Net neural network is used for fine semantic segmentation, and a particle swarm optimization algorithm and a convolutional neural network are combined for scene category prediction to construct a multi-scene classification and recognition method. Through Gaussian membership function and Q-reinforcement learning, combined with fuzzy logic, adaptive optimization of lighting parameters is carried out to construct a fuzzy rule base covering multiple scenes and user behaviors, and the parameter vector is updated using gradient descent.
It achieves accurate classification and recognition of complex indoor multi-scenes, significantly improves scene classification accuracy, accurately matches the objective lighting needs of different scenes, and caters to users' personalized preferences, thereby improving user comfort.
Smart Images
Figure CN122244654A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automatic control, specifically to a multi-scene automatic control system for lighting based on deep learning. Background Technology
[0002] Multi-scene automatic lighting control systems can automatically adjust parameters such as brightness, color, and color temperature of lights according to different scenarios, learning users' personalized habits and preferences to provide more personalized services. However, traditional scene recognition methods lack effective mining of spatial structure information in deep images, have coarse semantic segmentation granularity, and do not fully integrate information complexity and spatial structure characteristics-related features, resulting in high classification confusion rates. This leads to a mismatch between subsequent lighting control parameters and actual scene requirements, affecting the user experience. At the same time, they often adopt fixed parameter mapping modes, failing to fully consider users' personalized operating habits, resulting in weak adaptive adjustment capabilities. They struggle to balance objective scene requirements with users' personalized preferences, leading to low matching degrees between lighting parameters and actual user needs, and the control logic lacks flexibility and learning ability. Summary of the Invention
[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a multi-scene automatic lighting control system based on deep learning. It addresses the technical problems of traditional scene recognition methods, such as the lack of effective mining of spatial structure information from depth images, coarse semantic segmentation granularity, insufficient integration of information complexity and spatial structure characteristics, high classification confusion rates, and subsequent mismatch between lighting control parameters and actual scene requirements, thus affecting user experience. This solution employs a full-scene classification and recognition method, utilizing a U-Net neural network for fine semantic segmentation of depth images to obtain multi-semantic region segmented images. Information entropy is extracted to represent scene information complexity, fuzzy entropy is extracted to represent spatial structure characteristics, color features from color images are integrated, redundant and irrelevant features are removed using a particle swarm optimization algorithm, and scene category prediction is performed using a convolutional neural network. This achieves accurate classification and recognition of complex indoor multi-scene scenarios, significantly improving scene classification accuracy. This solution provides accurate and reliable scenario-based decision-making for adaptive lighting control. Addressing the technical issues of traditional lighting control methods that often employ fixed parameter mapping, failing to fully consider users' personalized operating habits, exhibiting weak adaptive adjustment capabilities, and struggling to balance objective scenario requirements with user preferences, resulting in low matching between lighting parameters and actual user needs, and a lack of flexibility and learning ability in the control logic, this solution uses a Gaussian membership function to fuzzify scene labels and user behavior logs, constructing a fuzzy rule library covering multiple scenarios and user behaviors. Through a zero-order TS inference engine, rule weights are normalized, and Q-reinforcement learning is integrated to convert user adjustments to the lighting into reward signals. Gradient descent is used to update the adjustable parameter vector of the fuzzy inference, achieving adaptive dynamic optimization of lighting parameters. This not only accurately matches the objective lighting needs of different scenarios but also aligns with user preferences, significantly improving user comfort.
[0004] The technical solution adopted by the present invention is as follows: The present invention provides a multi-scene automatic lighting control system based on deep learning, which includes a data acquisition module, an image preprocessing module, a scene classification module and a lighting control module.
[0005] The data acquisition module includes a scene image acquisition unit and a user behavior log recording unit. The scene image acquisition unit uses a color camera and depth image to acquire images of the current environment and records them through a polling method. The user behavior log recording unit records the time, type, and state information before and after the user operates on the lights, forming a user behavior log.
[0006] The image preprocessing module acquires the current environment image, which includes a color image with RGB channels, a corresponding depth image with a depth channel, and scene labels. A bilateral filtering algorithm is used to preprocess the depth image. Filtering is achieved through a weighted average of spatial distance weights and depth intensity difference weights, removing noise from the depth image while preserving key edges and structural features, resulting in a preprocessed depth image. The formula used is as follows: ;
[0007] In the formula, It is a preprocessed depth image. These are the pixel coordinates of the depth image. It is a normalization factor. It is a pixel. The set of neighboring pixels, It is any coordinate within the set of neighboring pixels. yes The original depth intensity value, It is the spatial standard deviation. It is the standard deviation of the depth range;
[0008] The scene classification module uses a full-scene classification and recognition method to perform semantic segmentation and recognition of objects in the current environment image and output scene labels.
[0009] The lighting control module uses an indoor adaptive lighting method that integrates Q-reinforcement learning and fuzzy logic to achieve adaptive control of the lighting fixtures.
[0010] Furthermore, in the scene classification module, the aforementioned full-scene classification and recognition method specifically includes the following steps:
[0011] Step A1: Establish and initialize a U-Net deep learning model. The U-Net deep learning model includes four parts: encoder, decoder, skip connections, and output layer. The preprocessed image is input into the encoder. The encoder adopts a 4-level downsampling structure, with each level containing 2 convolutional layers and 1 pooling layer. The decoder adopts a 4-level upsampling structure, with each level containing 1 upsampling layer, 1 feature fusion layer, and 2 convolutional layers. The output layer maps the feature map output by the decoder to obtain a semantic segmentation map.
[0012] Step A2: Based on the segmented semantic segmentation graph, extract two types of entropy base features: information entropy and fuzzy entropy, to characterize the information complexity and spatial structure characteristics of the scene. Specifically, this includes: counting the depth value of each semantic segmentation graph, obtaining the discrete distribution set of depth values, calculating the probability mass function of each discrete depth value, and calculating the information entropy value of the semantic segmentation graph.
[0013] The depth value of each semantic segmentation graph is normalized and mapped to the interval [0, 1] to obtain the normalized depth value. The fuzzy membership function is defined as a Sigmoid type membership function, and the fuzzy entropy value of the semantic segmentation graph is calculated.
[0014] Step A3: For each semantic segmentation map, fuse the information entropy, fuzzy entropy, and the color mean and color variance features of the corresponding color image to obtain the feature vector of the semantic segmentation map. Concatenate the feature vectors of all semantic segmentation maps to obtain the initial feature set of the current environment image.
[0015] Step A4: Use the particle swarm optimization algorithm to filter the initial feature set, remove redundant and irrelevant features, and obtain the optimal feature combination;
[0016] Step A5: Build and initialize a convolutional neural network as a scene classification model, input the optimal feature combination into the scene classification model, perform scene category prediction, and output the final scene label.
[0017] Furthermore, in the lighting control module, the aforementioned indoor adaptive lighting method specifically includes the following steps:
[0018] Step B1: Use a Gaussian membership function to fuzzify the scene labels and user behavior logs, construct a fuzzy rule base, and use a zero-order TS inference engine to normalize the rule weights in the fuzzy rule base to obtain the weighted average of all rule outputs. The formula used is as follows: ;
[0019] In the formula, yes The output values of the lighting parameters at any given time. yes Moment-time scene tags and user behavior logs, It is the number of rules in the fuzzy rule base. Yes Traversal, These are the normalized rule weights. These are the corresponding rule output parameters;
[0020] Step B2: Using Q-reinforcement learning combined with user behavior habits, adjust the lighting parameters. Establish a Q-table to store scene labels and the expected cumulative reward values for the lighting parameters. Convert user adjustments to the lighting parameters into reward signals. Design a reward function using the following formula: ;
[0021] In the formula, yes Reward value at any moment It is the penalty coefficient. These are the preferred lighting parameters confirmed through user behavior log feedback;
[0022] Step B3: Update the adjustable parameter vector of fuzzy inference using gradient descent, define the objective function as temporal difference error square, obtain the final lighting parameters through Q reinforcement learning, and adjust them through the built-in microcontroller of the lighting fixture.
[0023] The beneficial effects achieved by the present invention using the above solution are as follows:
[0024] (1) In view of the technical problems in traditional scene recognition methods, such as lack of effective mining of spatial structure information of depth images, coarse semantic segmentation, insufficient integration of information complexity and spatial structure characteristics, high classification confusion rate, which makes the subsequent lighting control parameters mismatch with actual scene requirements and affects user experience, this solution uses a full scene classification and recognition method to perform fine semantic segmentation of depth images using U-Net neural network to obtain multi-semantic region segmentation images, extract information entropy to represent scene information complexity, extract fuzzy entropy to represent spatial structure characteristics, integrate color features of color images, remove redundant and irrelevant features through particle swarm optimization algorithm, and perform scene category prediction through convolutional neural network to achieve accurate classification and recognition of complex indoor multi-scenes, significantly improve scene classification accuracy, and provide accurate and reliable scene decision basis for adaptive lighting control;
[0025] (2) In view of the technical problems that traditional lighting control adopts fixed parameter mapping mode, does not fully consider the user's personalized operation habits, has weak adaptive adjustment capability, and is difficult to take into account the objective needs of the scene and the user's personalized preferences, resulting in low matching degree between lighting parameters and actual user needs, and lack of flexibility and learning ability of control logic, this solution uses Gaussian membership function to fuzzify scene labels and user behavior logs, builds a fuzzy rule library covering multiple scenes and multiple user behaviors, normalizes the rule weights through zero-order TS inference engine, integrates Q reinforcement learning, converts the user's adjustment operation of lighting into reward signal, and uses gradient descent method to update the fuzzy inference adjustable parameter vector to realize adaptive dynamic optimization of lighting parameters, which can accurately match the objective lighting needs of different scenes and fit the user's personalized preferences, significantly improving the user's comfort. Attached Figure Description
[0026] Figure 1 The present invention provides a module connection diagram for a multi-scene automatic lighting control system based on deep learning.
[0027] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0029] Example 1, see Figure 1 The present invention provides a multi-scene automatic lighting control system based on deep learning, which includes a data acquisition module, an image preprocessing module, a scene classification module and a lighting control module.
[0030] The data acquisition module includes a scene image acquisition unit and a user behavior log recording unit. The scene image acquisition unit uses a color camera and depth image to acquire images of the current environment and records them through a polling method. The user behavior log recording unit records the time, type, and state information before and after the user operates on the lights, forming a user behavior log.
[0031] The image preprocessing module acquires the current environment image, which includes a color image with RGB channels, a corresponding depth image with a depth channel, and scene labels. A bilateral filtering algorithm is used to preprocess the depth image. Filtering is achieved through a weighted average of spatial distance weights and depth intensity difference weights, removing noise from the depth image while preserving key edges and structural features, resulting in a preprocessed depth image. The formula used is as follows: ;
[0032] In the formula, It is a preprocessed depth image. These are the pixel coordinates of the depth image. It is a normalization factor. It is a pixel. The set of neighboring pixels, It is any coordinate within the set of neighboring pixels. yes The original depth intensity value, It is the spatial standard deviation. It is the standard deviation of the depth range;
[0033] The scene classification module uses a full-scene classification and recognition method to perform semantic segmentation and recognition of objects in the current environment image and output scene labels.
[0034] The lighting control module uses an indoor adaptive lighting method that integrates Q-reinforcement learning and fuzzy logic to achieve adaptive control of the lighting fixtures.
[0035] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the scene classification module, the full-scene classification and recognition method specifically includes the following steps:
[0036] Step A1: Establish and initialize a U-Net deep learning model. The U-Net deep learning model includes four parts: encoder, decoder, skip connections, and output layer. The preprocessed image is input into the encoder. The encoder adopts a 4-level downsampling structure, with each level containing 2 convolutional layers and 1 pooling layer. The decoder adopts a 4-level upsampling structure, with each level containing 1 upsampling layer, 1 feature fusion layer, and 2 convolutional layers. The output layer maps the feature map output by the decoder to obtain a semantic segmentation map.
[0037] Step A2: Based on the segmented semantic segmentation graph, extract two types of entropy base features: information entropy and fuzzy entropy, to characterize the information complexity and spatial structure characteristics of the scene. Specifically, this includes: counting the depth value of each semantic segmentation graph, obtaining the discrete distribution set of depth values, calculating the probability mass function of each discrete depth value, and calculating the information entropy value of the semantic segmentation graph.
[0038] The depth value of each semantic segmentation graph is normalized and mapped to the interval [0, 1] to obtain the normalized depth value. The fuzzy membership function is defined as a Sigmoid type membership function, and the fuzzy entropy value of the semantic segmentation graph is calculated.
[0039] Step A3: For each semantic segmentation map, fuse the information entropy, fuzzy entropy, and the color mean and color variance features of the corresponding color image to obtain the feature vector of the semantic segmentation map. Concatenate the feature vectors of all semantic segmentation maps to obtain the initial feature set of the current environment image.
[0040] Step A4: Use the particle swarm optimization algorithm to filter the initial feature set, remove redundant and irrelevant features, and obtain the optimal feature combination;
[0041] Step A5: Build and initialize a convolutional neural network as a scene classification model, input the optimal feature combination into the scene classification model, perform scene category prediction, and output the final scene label.
[0042] Example 3, see Figure 1 This embodiment is based on the above embodiment. In the lighting control module, the indoor adaptive lighting method specifically includes the following steps:
[0043] Step B1: Use a Gaussian membership function to fuzzify the scene labels and user behavior logs, construct a fuzzy rule base, and use a zero-order TS inference engine to normalize the rule weights in the fuzzy rule base to obtain the weighted average of all rule outputs. The formula used is as follows: ;
[0044] In the formula, yes The output values of the lighting parameters at any given time. yes Moment-time scene tags and user behavior logs, It is the number of rules in the fuzzy rule base. Yes Traversal, These are the normalized rule weights. These are the corresponding rule output parameters;
[0045] Step B2: Using Q-reinforcement learning combined with user behavior habits, adjust the lighting parameters. Establish a Q-table to store scene labels and the expected cumulative reward values for the lighting parameters. Convert user adjustments to the lighting parameters into reward signals. Design a reward function using the following formula: ;
[0046] In the formula, yes Reward value at any moment It is the penalty coefficient. These are the preferred lighting parameters confirmed through user behavior log feedback;
[0047] Step B3: Update the adjustable parameter vector of fuzzy inference using gradient descent, define the objective function as temporal difference error square, obtain the final lighting parameters through Q reinforcement learning, and adjust them through the built-in microcontroller of the lighting fixture.
[0048] Example 4 is based on the above examples. In Example 2, the encoder adopts a 4-level downsampling structure, each level containing 2 convolutional layers and 1 pooling layer. The convolutional layers use 3×3 convolutional kernels with a stride of 1 and a SAME padding method. Each convolutional layer is followed by a ReLU activation function. The pooling layer uses a 2×2 average pooling window with a stride of 2.
[0049] The decoder adopts a 4-level upsampling structure, each level containing one upsampling layer, one feature fusion layer and two convolutional layers. Its upsampling layer adopts transposed convolution operation, with a convolution kernel size of 2×2 and a stride of 2. The convolutional layers are the same as those of the encoder.
[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
[0052] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A multi-scene automatic lighting control system based on deep learning, characterized in that, It includes a data acquisition module, an image preprocessing module, a scene classification module, and a lighting control module; The data acquisition module includes a scene image acquisition unit and a user behavior log recording unit. The scene image acquisition unit uses a color camera and depth image to acquire images of the current environment, and the user behavior log recording unit records user behavior logs when the user operates the lights. The image preprocessing module acquires the current environment image, which includes a color image with RGB channels, a corresponding depth image with Depth channels, and scene labels. The depth image is preprocessed using a bilateral filtering algorithm to obtain a preprocessed depth image. The scene classification module uses a full-scene classification and recognition method to perform semantic segmentation and recognition of objects in the current environment image and output scene labels. The lighting control module uses an indoor adaptive lighting method that integrates Q-reinforcement learning and fuzzy logic to achieve adaptive control of the lighting fixtures.
2. The multi-scene automatic lighting control system based on deep learning according to claim 1, characterized in that, In the scene classification module, the aforementioned full-scene classification and recognition method specifically includes the following steps: Step A1: Establish and initialize a U-Net deep learning model. The U-Net deep learning model includes four parts: encoder, decoder, skip connections, and output layer. The preprocessed image is input into the encoder, and the output layer maps the feature map output by the decoder to obtain a semantic segmentation map. Step A2: Based on the segmented semantic segmentation map, extract two types of entropy base features: information entropy and fuzzy entropy. Specifically, this includes: counting the depth value of each semantic segmentation map, obtaining the discrete distribution set of the depth values, calculating the probability mass function of each discrete depth value, and calculating the information entropy value of the semantic segmentation map. The depth value of each semantic segmentation graph is normalized and mapped to the interval [0, 1] to obtain the normalized depth value. The fuzzy membership function is defined as a Sigmoid type membership function, and the fuzzy entropy value of the semantic segmentation graph is calculated. Step A3: For each semantic segmentation map, fuse the information entropy, fuzzy entropy, and the color mean and color variance features of the corresponding color image to obtain the feature vector of the semantic segmentation map. Concatenate the feature vectors of all semantic segmentation maps to obtain the initial feature set of the current environment image. Step A4: Use the particle swarm optimization algorithm to filter the initial feature set, remove redundant and irrelevant features, and obtain the optimal feature combination; Step A5: Build and initialize a convolutional neural network as a scene classification model, input the optimal feature combination into the scene classification model, perform scene category prediction, and output the final scene label.
3. The multi-scene automatic lighting control system based on deep learning according to claim 1, characterized in that, In the lighting control module, the aforementioned indoor adaptive lighting method specifically includes the following steps: Step B1: Use the Gaussian membership function to fuzzify the scene labels and user behavior logs, build a fuzzy rule base, and use the zero-order TS inference engine to normalize the rule weights in the fuzzy rule base to obtain the weighted average of all rule outputs. Step B2: Adjust the lighting parameters by combining Q reinforcement learning with user behavior habits, establish a Q table to store the scene labels and the expected cumulative reward values of the lighting parameters, convert the user's adjustment of the lighting parameters into reward signals, and design a reward function; Step B3: Update the adjustable parameter vector of fuzzy inference using gradient descent, define the objective function as temporal difference error square, obtain the final lighting parameters through Q reinforcement learning, and adjust them through the built-in microcontroller of the lighting fixture.