Gamma debugging system and debugging method based on deep learning
By using the Gamma debugging system based on deep learning, the problems of differences in display device hardware characteristics and insufficient adaptability to dynamic scenes have been solved, achieving efficient and stable display effects and low power consumption, thereby improving user experience and production efficiency.
Patent Information
- Application Number
- CN202512011813.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-29
- Publication Date
- 2026-02-24
AI Technical Summary
Existing technologies cannot adapt to the differences in hardware characteristics of different display devices, have insufficient adaptability to dynamic scenes, low debugging efficiency, and insufficient real-time inference capabilities. Furthermore, traditional display driver chips lack hardware support for AI algorithms, resulting in poor display effects and high debugging costs.
A deep learning-based Gamma tuning system is adopted, which includes a data acquisition layer, an algorithm processing layer, and a hardware execution layer. It collects data through environmental sensors and image sensors, extracts features from the displayed images using a deep learning prediction engine, predicts the optimal Gamma lookup table parameters, and dynamically updates and compensates for the effects of display panel temperature through the hardware execution layer.
It achieves improved cross-device Gamma consistency, enhanced display stability under varying ambient light conditions, significantly reduced debugging time, lower power consumption, and markedly improved display quality and user experience, while also reducing manual debugging costs and production inconsistencies.
Smart Images
Figure CN121565077A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to Gamma debugging, specifically to a deep learning-based Gamma debugging system and method. Background Technology
[0002] In the field of display technology, gamma correction is a crucial step that directly affects the brightness distribution, contrast, and color reproduction of an image. Traditional gamma adjustment methods mainly include:
[0003] 1) Fixed Gamma Curve Method: Use a standard Gamma curve (such as 2.2 of sRGB standard, 2.2 / 2.6 of Adobe RGB), store the predefined Gamma parameters through a lookup table (LUT), and fix them as fixed parameters in the display driver chip;
[0004] 2) Manual debugging method: Relying on the experience of engineers, manual debugging is carried out using an oscilloscope and colorimeter to adjust the Gamma voltage value and lookup table parameters, and the optimal parameters are determined through visual inspection and objective test data.
[0005] The existing technology mainly has the following technical problems:
[0006] 1) Device Differences: Different display devices (such as LCD, OLED, Micro-OLED, etc.) have significant differences in physical characteristics. The display effect of the same image on different display devices varies to a certain extent. Traditional methods cannot adapt to the differences in hardware characteristics.
[0007] 2) Insufficient adaptability to dynamic scenes: The fixed Gamma curve method cannot respond to changes in ambient light. Different image content requires different Gamma correction strategies. HDR (High Dynamic Range) content has higher requirements for Gamma accuracy. Real-time scene changes lead to poor display effects.
[0008] 3) Low debugging efficiency: Manual debugging methods rely on expert experience, which is time-consuming and costly. It is difficult to guarantee consistency in large-scale production, and the debugging accuracy is greatly affected by subjective factors.
[0009] 4) Hardware limitations: Traditional display driver chips lack hardware support for AI algorithms, have insufficient real-time inference capabilities, insufficient neural network inference computing power, complex register configurations, and slow parameter update speeds. Summary of the Invention
[0010] (a) Technical problems to be solved
[0011] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a deep learning-based Gamma debugging system and debugging method, which can effectively overcome the shortcomings of the existing technology in that it is difficult to debug the Gamma parameter accurately and efficiently, and the lack of adaptability to dynamic scenes.
[0012] (II) Technical Solution
[0013] To achieve the above objectives, the present invention provides the following technical solution:
[0014] A deep learning-based Gamma debugging system includes a data acquisition layer, an algorithm processing layer, and a hardware execution layer.
[0015] The data acquisition layer collects environmental parameters, displayed images from display devices, and the operating temperature of the display panel, and stores the physical characteristic parameters of different display devices and user preference settings for different scenarios;
[0016] The algorithm processing layer extracts image features from the displayed image, encodes environmental parameters into environmental features, predicts the optimal Gamma lookup table parameters based on the image features and environmental features, and performs nonlinear compensation based on the physical characteristic parameters of the display device.
[0017] The hardware execution layer deploys the algorithm processing layer, dynamically updates the Gamma lookup table parameters, and compensates for the impact of the display panel's operating temperature on the Gamma lookup table parameters.
[0018] Preferably, the data acquisition layer includes environmental sensors, image sensors, and a historical database;
[0019] An environmental sensor detects environmental parameters in real time, including ambient light intensity and screen brightness, as well as the operating temperature of the display panel.
[0020] Image sensors acquire images displayed on display devices in real time;
[0021] The historical database includes a device parameter database and a user preference database. The device parameter database stores the physical characteristic parameters of different display devices, while the user preference database stores the user's preference settings for different scenarios.
[0022] Preferably, the algorithm processing layer includes a deep learning prediction engine and a dynamic compensation module;
[0023] The deep learning prediction engine extracts image features from the displayed image and encodes environmental parameters into environmental features. Based on the image features and environmental features, it predicts the optimal Gamma lookup table parameters.
[0024] The dynamic compensation module performs nonlinear compensation based on the physical characteristic parameters of the display device.
[0025] Preferably, the deep learning prediction engine includes an input layer, a feature extraction layer, a temporal processing layer, a feature fusion layer, and an output layer;
[0026] The input layer receives the displayed image and environmental parameters.
[0027] The feature extraction layer includes an image feature extraction module, which uses a convolutional neural network (CNN) to extract image features, including RGB data, image brightness, image color, and image texture, from the displayed image.
[0028] The temporal processing layer includes an environmental parameter encoding module, which uses an LSTM network to encode environmental parameters into environmental features.
[0029] The feature fusion layer uses an attention mechanism and feature concatenation operation to weightedly fuse image features and environmental features to obtain fused features;
[0030] The output layer includes a Gamma prediction module and a post-processing module. The Gamma prediction module uses a fully connected network to output the optimal Gamma lookup table parameters based on the fused features. The post-processing module is used to perform parameter normalization and boundary constraints.
[0031] Preferably, the image feature extraction module uses ResNet-50 as the backbone network of a CNN, and ResNet-50 includes convolutional layers, residual blocks, and global average pooling layers;
[0032] The convolutional layer includes a first convolutional layer Conv1 and a second convolutional layer Conv2. The first convolutional layer Conv1 includes 64 3*3 convolutional kernels and uses the ReLU activation function.
[0033] The residual blocks, including the second residual block Res2, the third residual block Res3, the fourth residual block Res4, and the fifth residual block Res5, form the core structure of ResNet-50.
[0034] The global average pooling layer, located at the end of ResNet-50, reduces the multidimensional image feature map to a 1024-dimensional image feature vector.
[0035] The LSTM network in the environmental parameter encoding module includes a first LSTM layer and a second LSTM layer. Both the first LSTM layer and the second LSTM layer include 256 hidden units for processing time-series information.
[0036] The fully connected network in the Gamma prediction module includes a first fully connected layer and a second fully connected layer, each containing 256 neurons.
[0037] Preferably, the image feature extraction module uses a convolutional neural network (CNN) to extract a 1024-dimensional image feature vector from the displayed image, including RGB data, image brightness, image color, and image texture.
[0038] The environmental parameter encoding module uses an LSTM network to encode environmental parameters into a 128-dimensional environmental feature vector.
[0039] The Gamma prediction module uses a fully connected network to output 256-dimensional optimal Gamma lookup table parameters based on the fusion features.
[0040] Preferably, the training method for the deep learning prediction engine includes:
[0041] S11. Collect 10,000 high-quality images containing different scenes, obtain the environmental parameters of each image, and obtain the standard Gamma parameters of each image through professional calibration to construct a historical dataset.
[0042] S12. Divide the historical dataset into training set, validation set and test set according to the preset ratio;
[0043] S13. Set the loss function and optimizer for the deep learning prediction engine;
[0044] S14. Input the training set into the deep learning prediction engine to train the model;
[0045] S15. Calculate the loss value based on the loss function, and the optimizer updates the model parameters based on the loss value and network gradient information;
[0046] S16. If the loss value is less than the preset threshold, the model training ends and the current deep learning prediction engine is the trained deep learning prediction engine; otherwise, return to S14 and continue to train the model using the training set.
[0047] S17. Input the validation set into the trained deep learning prediction engine, evaluate the model's generalization ability by observing its performance on the validation set, and fine-tune the model's hyperparameters and structure.
[0048] S18. Input the test set into the optimized deep learning prediction engine to evaluate the model performance.
[0049] Preferably, the loss function is:
[0050] Loss function = content loss + perceptual loss + adversarial loss;
[0051] Among them, content loss is pixel-level error based on PSNR and MSE, perceptual loss is perceptual similarity loss based on VGG network, and adversarial loss is loss from image quality adversarial training based on GAN network.
[0052] Preferably, the hardware execution layer includes a display driver chip, a register configuration module, a temperature compensation module, and a display device;
[0053] The display driver chip integrates an AI acceleration unit, deploys a deep learning prediction engine and a dynamic compensation module, and supports real-time deep learning inference.
[0054] Register configuration module, dynamically updates Gamma lookup table parameters;
[0055] The temperature compensation module compensates for the impact of the display panel's operating temperature on the Gamma lookup table parameters.
[0056] A deep learning-based Gamma debugging method includes the following steps:
[0057] S1. Data Acquisition: Acquire the current display image of the display device, as well as the corresponding environmental parameters, and analyze the brightness histogram of the display image;
[0058] S2. Feature Extraction: Extract image features, including RGB data, image brightness, image color, and image texture, from the displayed image, encode environmental parameters into environmental features, and calculate the statistical features of the displayed image.
[0059] S3, Gamma Prediction: Based on the attention mechanism and feature concatenation operation, image features and environmental features are weighted and fused to obtain fused features. The optimal Gamma lookup table parameters are output based on the fused features.
[0060] S4. Parameter Verification: Smooth the parameters of the optimal Gamma lookup table, verify their rationality, check whether they exceed hardware limitations, and perform boundary constraint processing.
[0061] S5. Hardware Update: Dynamically update the parameters of the Gamma lookup table, configure them, and verify the update results;
[0062] S6. Effect Evaluation: Calculate the adjusted display image quality, record historical adjustment data, and provide feedback for model optimization.
[0063] (III) Beneficial Effects
[0064] Compared with existing technologies, the deep learning-based Gamma debugging system and method provided by this invention have the following advantages:
[0065] 1) Technical effects
[0066] Improved display consistency: Cross-device Gamma consistency reaches over 95%, display stability under ambient light changes is improved by 60%, and the display quality of HDR content is significantly improved;
[0067] Improved debugging efficiency: Debugging time is reduced from 2-4 hours to 5 minutes, automation level reaches over 90%, and consistency in large-scale production is improved by 85%;
[0068] Power consumption optimization: Average power consumption is reduced by 8%~15%, and the battery life of mobile devices is extended by 10%~20%, reducing unnecessary power consumption through intelligent adjustment;
[0069] 2) Commercial value
[0070] Cost reduction: Reduce manual debugging costs by 70%, increase production efficiency by 50%, and reduce quality control costs;
[0071] Product competitiveness: Significantly improved display quality, noticeably enhanced user experience, and a competitive advantage through technological differentiation;
[0072] Market Applications: Applicable to multiple fields such as mobile phones, tablets, TVs, and automotive displays; it has broad application prospects in emerging markets such as AR / VR and smartphones; and it has high application value in professional fields such as industrial displays and professional displays.
[0073] 3) Social benefits
[0074] Energy saving and environmental protection: Reduced overall power consumption reduces carbon emissions, extends the lifespan of display devices, and reduces electronic waste generation;
[0075] User experience: Visual comfort is significantly improved, reducing visual fatigue and the risk of myopia, and improving the user's quality of life. Attached Figure Description
[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0077] Figure 1 This is a schematic diagram of the system of the present invention;
[0078] Figure 2 This is a schematic diagram of the deep learning prediction engine in this invention;
[0079] Figure 3 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0080] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.
[0081] The following describes the specific architecture of a deep learning-based Gamma debugging system provided by this invention, using concrete examples (e.g.) Figure 1 (As shown) and its technical effects. The system architecture includes: a data acquisition layer, an algorithm processing layer, and a hardware execution layer;
[0082] The data acquisition layer collects environmental parameters, displayed images from display devices, and the operating temperature of the display panel, and stores the physical characteristic parameters of different display devices and user preference settings for different scenarios;
[0083] The algorithm processing layer extracts image features from the displayed image, encodes environmental parameters into environmental features, predicts the optimal Gamma lookup table parameters based on the image features and environmental features, and performs nonlinear compensation based on the physical characteristic parameters of the display device.
[0084] The hardware execution layer deploys the algorithm processing layer, dynamically updates the Gamma lookup table parameters, and compensates for the impact of the display panel's operating temperature on the Gamma lookup table parameters.
[0085] I. Data Acquisition Layer
[0086] The data acquisition layer includes environmental sensors, image sensors, and a historical database;
[0087] An environmental sensor detects environmental parameters in real time, including ambient light intensity (range 0~30000 lux, accuracy ±5%) and screen brightness, as well as the operating temperature of the display panel (accuracy ±1°C).
[0088] Image sensor, which captures real-time images of the display device (1920*1080 resolution, 60Hz refresh rate);
[0089] The historical database includes a device parameter database and a user preference database. The device parameter database stores the physical characteristic parameters of different display devices, while the user preference database stores the user's preference settings for different scenarios.
[0090] II. Algorithm Processing Layer
[0091] The algorithm processing layer includes a deep learning prediction engine and a dynamic compensation module;
[0092] The deep learning prediction engine extracts image features from the displayed image and encodes environmental parameters into environmental features. Based on the image features and environmental features, it predicts the optimal Gamma lookup table parameters.
[0093] The dynamic compensation module performs nonlinear compensation based on the physical characteristic parameters of the display device.
[0094] like Figure 2 As shown, the deep learning prediction engine includes an input layer, a feature extraction layer, a temporal processing layer, a feature fusion layer, and an output layer.
[0095] The input layer receives the displayed image and environmental parameters.
[0096] The feature extraction layer includes an image feature extraction module, which uses a convolutional neural network (CNN) to extract image features, including RGB data, image brightness, image color, and image texture, from the displayed image.
[0097] The timing processing layer includes an environment parameter encoding module, which uses an LSTM network (or Transformer network) to encode environment parameters into environment features;
[0098] The feature fusion layer uses an attention mechanism and feature concatenation operation to weightedly fuse image features and environmental features to obtain fused features;
[0099] The output layer includes a Gamma prediction module and a post-processing module. The Gamma prediction module uses a fully connected network to output the optimal Gamma lookup table parameters based on the fused features. The post-processing module is used to perform parameter normalization and boundary constraints.
[0100] Specifically, the image feature extraction module uses ResNet-50 as the backbone network of the CNN. ResNet-50 includes convolutional layers, residual blocks, and global average pooling layers.
[0101] The convolutional layer includes a first convolutional layer Conv1 and a second convolutional layer Conv2. The first convolutional layer Conv1 includes 64 3*3 convolutional kernels and uses the ReLU activation function.
[0102] The residual blocks, including the second residual block Res2, the third residual block Res3, the fourth residual block Res4, and the fifth residual block Res5, form the core structure of ResNet-50.
[0103] The global average pooling layer, located at the end of ResNet-50, reduces the multidimensional image feature map to a 1024-dimensional image feature vector.
[0104] The LSTM network in the environmental parameter encoding module includes a first LSTM layer and a second LSTM layer. Both the first and second LSTM layers contain 256 hidden units for processing timing information.
[0105] The fully connected network in the Gamma prediction module includes a first fully connected layer and a second fully connected layer, each containing 256 neurons.
[0106] Specifically, the image feature extraction module uses a convolutional neural network (CNN) to extract a 1024-dimensional image feature vector from the displayed image, including RGB data, image brightness, image color, and image texture.
[0107] The environmental parameter encoding module uses an LSTM network to encode environmental parameters into 128-dimensional environmental feature vectors.
[0108] The Gamma prediction module uses a fully connected network to output 256-dimensional optimal Gamma lookup table parameters based on fused features.
[0109] The training method for the deep learning prediction engine in this application includes:
[0110] S11. Collect 10,000 high-quality images containing different scenes, obtain the environmental parameters of each image, and obtain the standard Gamma parameters of each image through professional calibration to construct a historical dataset.
[0111] S12. Divide the historical dataset into training set, validation set and test set according to the preset ratio;
[0112] S13. Set the loss function and optimizer for the deep learning prediction engine;
[0113] S14. Input the training set into the deep learning prediction engine to train the model;
[0114] S15. Calculate the loss value based on the loss function, and the optimizer updates the model parameters based on the loss value and network gradient information;
[0115] S16. If the loss value is less than the preset threshold, the model training ends and the current deep learning prediction engine is the trained deep learning prediction engine; otherwise, return to S14 and continue to train the model using the training set.
[0116] S17. Input the validation set into the trained deep learning prediction engine, evaluate the model's generalization ability by observing its performance on the validation set, and fine-tune the model's hyperparameters and structure.
[0117] S18. Input the test set into the optimized deep learning prediction engine to evaluate the model performance.
[0118] Specifically, the loss function is:
[0119] Loss function = content loss + perceptual loss + adversarial loss;
[0120] Among them, content loss is pixel-level error based on PSNR and MSE, perceptual loss is perceptual similarity loss based on VGG network, and adversarial loss is loss from image quality adversarial training based on GAN network.
[0121] III. Hardware Execution Layer
[0122] The hardware execution layer includes a display driver chip, a register configuration module, a temperature compensation module, and a display device;
[0123] The display driver chip integrates an AI acceleration unit (TOPS≥1), deploys a deep learning prediction engine and a dynamic compensation module, and supports real-time deep learning inference.
[0124] Register configuration module, dynamically updates Gamma lookup table parameters;
[0125] The temperature compensation module compensates for the impact of the display panel's operating temperature on the Gamma lookup table parameters.
[0126] Based on the aforementioned deep learning-based Gamma debugging system, this application also discloses a deep learning-based Gamma debugging method, such as... Figure 3 As shown, it includes the following steps:
[0127] S1. Data Acquisition: Acquire the current display image of the display device, as well as the corresponding environmental parameters, and analyze the brightness histogram of the display image;
[0128] S2. Feature Extraction: Extract image features, including RGB data, image brightness, image color, and image texture, from the displayed image, encode environmental parameters into environmental features, and calculate the statistical features of the displayed image.
[0129] S3, Gamma Prediction: Based on the attention mechanism and feature concatenation operation, image features and environmental features are weighted and fused to obtain fused features. The optimal Gamma lookup table parameters are output based on the fused features.
[0130] S4. Parameter Verification: Smooth the parameters of the optimal Gamma lookup table, verify their rationality, check whether they exceed hardware limitations, and perform boundary constraint processing.
[0131] S5. Hardware Update: Dynamically update the parameters of the Gamma lookup table, configure them, and verify the update results;
[0132] S6. Effect Evaluation: Calculate the adjusted display image quality, record historical adjustment data, and provide feedback for model optimization.
[0133] To better illustrate the technical solution of this application, three specific examples are provided below for detailed explanation.
[0134] Example 1: Adaptive Gamma Adjustment System for LCD Displays
[0135] A. System Configuration
[0136] Processor: ARM Cortex-A7 @ 1.2GHz + AI acceleration unit (0.5TOPS);
[0137] Storage: 512KB Flash + 2MB RAM;
[0138] Sensors: Ambient light sensor (APDS-9960), image sensor;
[0139] Monitor: 10.1-inch LCD panel, 1920*1200 resolution.
[0140] B. Implementation Steps
[0141] Step 1: System Initialization
[0142] 1) Load the parameters of the pre-trained deep learning model;
[0143] 2) Initialize the ambient light sensor and temperature sensor;
[0144] 3) Configure the initial value of the Gamma adjustment register;
[0145] Step 2: Data Acquisition and Processing
[0146] 1) The ambient light sensor collects ambient light intensity data every 100ms;
[0147] 2) The image processing unit analyzes the brightness distribution of the currently displayed image;
[0148] 3) Calculate the statistical characteristics of the image, such as histogram, contrast, and saturation;
[0149] Step 3: Gamma parameter prediction
[0150] 1) Input image features and environmental parameters into the deep learning model;
[0151] 2) The model outputs a 256-dimensional Gamma lookup table of parameters;
[0152] 3) Smooth the prediction results and apply boundary constraints;
[0153] Step 4: Hardware Parameter Update
[0154] 1) Write the Gamma parameter into the lookup table register of the display driver chip;
[0155] 2) Verify that the parameters were written successfully and update the system status;
[0156] 3) Record the adjustment history to provide data for subsequent optimization.
[0157] C. Verification of debugging results
[0158] Indoor environment (300 lux): Gamma value optimized from 2.2 to 2.1, improving details in dark areas;
[0159] Outdoor environment (5000 lux): Gamma value adjusted to 2.3 to enhance contrast - Response time: Parameter update latency less than 50ms - Energy consumption optimization: 8% energy saving compared to fixed Gamma.
[0160] Example 2: OLED Display HDR Content Optimization System
[0161] A. System Configuration
[0162] Processor: ARM Cortex-A53 @ 1.8GHz + AI acceleration unit (2TOPS);
[0163] Storage: 2MB Flash + 8MB RAM;
[0164] Sensors: High-precision ambient light sensor, image sensor;
[0165] Display: 6.7-inch OLED panel, supports HDR10+.
[0166] B. Special processing of HDR content:
[0167] Step 1: HDR Content Detection
[0168] 1) Parse HDR metadata to obtain brightness range and color gamut information;
[0169] 2) Check if the current content is in HDR format;
[0170] 3) Analyze the dynamic range and peak brightness of the content;
[0171] Step 2: Multi-level Gamma Correction
[0172] 1) For dark areas (0~0.3): Use a smaller Gamma value (1.8~2.0);
[0173] 2) For midtones (0.3~0.7): Use the standard Gamma value (2.2);
[0174] 3) For highlights (0.7~1.0): Use a larger Gamma value (2.4~2.6);
[0175] Step 3: Color Management
[0176] 1) Adjust the RGB weights according to the color gamut coverage of the display device;
[0177] 2) Adaptive compensation for saturation;
[0178] 3) Optimize contrast while maintaining color accuracy.
[0179] C. Optimization effect
[0180] HDR content display is significantly improved;
[0181] Dark detail improved by 30%;
[0182] Highlight clipping reduced by 25%;
[0183] Color accuracy remains above 90%.
[0184] Example 3: Micro-OLED AR / VR Display System
[0185] A. System Configuration
[0186] Processor: Qualcomm XR2 chip + AI acceleration unit (4TOPS);
[0187] Storage: 4MB Flash + 16MB RAM;
[0188] Sensors: Dual ambient light sensors, eye-tracking sensor;
[0189] Display: Dual Micro-OLED displays, 3200*3200 resolution.
[0190] B. Handling Special Needs of AR / VR
[0191] Step 1: User Status Awareness
[0192] 1) Eye-tracking data is used to obtain the user's gaze area;
[0193] 2) Detect the user's visual fatigue state;
[0194] 3) Analyze user adaptability to different lighting environments;
[0195] Step 2: Personalized Gamma Adjustment
[0196] 1) Adjust the Gamma curve according to user preferences;
[0197] 2) Perform high-precision Gamma adjustment on the gaze area;
[0198] 3) Energy-saving mode is adopted in the edge area;
[0199] Step 3: Real-time optimization
[0200] 1) Update the Gamma parameter of the gaze region in each frame;
[0201] 2) Adjust the brightness limit according to the user's level of fatigue;
[0202] 3) Dynamically balance display effect and power consumption.
[0203] C. Performance Indicators:
[0204] Response time: <16ms (60Hz refresh rate);
[0205] User experience rating: improved by 40%;
[0206] Battery life: Extended by 20%;
[0207] Visual comfort: Reduces eye fatigue by 25%.
[0208] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A deep learning-based Gamma debugging system, characterized in that: It includes a data acquisition layer, an algorithm processing layer, and a hardware execution layer; The data acquisition layer collects environmental parameters, displayed images from display devices, and the operating temperature of the display panel, and stores the physical characteristic parameters of different display devices and user preference settings for different scenarios; The algorithm processing layer extracts image features from the displayed image, encodes environmental parameters into environmental features, predicts the optimal Gamma lookup table parameters based on the image features and environmental features, and performs nonlinear compensation based on the physical characteristic parameters of the display device. The hardware execution layer deploys the algorithm processing layer, dynamically updates the Gamma lookup table parameters, and compensates for the impact of the display panel's operating temperature on the Gamma lookup table parameters.
2. The deep learning-based Gamma debugging system according to claim 1, characterized in that: The data acquisition layer includes environmental sensors, image sensors, and a historical database; An environmental sensor detects environmental parameters in real time, including ambient light intensity and screen brightness, as well as the operating temperature of the display panel. Image sensors acquire images displayed on display devices in real time; The historical database includes a device parameter database and a user preference database. The device parameter database stores the physical characteristic parameters of different display devices, while the user preference database stores the user's preference settings for different scenarios.
3. The deep learning-based Gamma debugging system according to claim 1, characterized in that: The algorithm processing layer includes a deep learning prediction engine and a dynamic compensation module; The deep learning prediction engine extracts image features from the displayed image and encodes environmental parameters into environmental features. Based on the image features and environmental features, it predicts the optimal Gamma lookup table parameters. The dynamic compensation module performs nonlinear compensation based on the physical characteristic parameters of the display device.
4. The deep learning-based Gamma debugging system according to claim 3, characterized in that: The deep learning prediction engine includes an input layer, a feature extraction layer, a temporal processing layer, a feature fusion layer, and an output layer; The input layer receives the displayed image and environmental parameters. The feature extraction layer includes an image feature extraction module, which uses a convolutional neural network (CNN) to extract image features, including RGB data, image brightness, image color, and image texture, from the displayed image. The temporal processing layer includes an environmental parameter encoding module, which uses an LSTM network to encode environmental parameters into environmental features. The feature fusion layer uses an attention mechanism and feature concatenation operation to weightedly fuse image features and environmental features to obtain fused features; The output layer includes a Gamma prediction module and a post-processing module. The Gamma prediction module uses a fully connected network to output the optimal Gamma lookup table parameters based on the fused features. The post-processing module is used to perform parameter normalization and boundary constraints.
5. The deep learning-based Gamma debugging system according to claim 4, characterized in that: The image feature extraction module uses ResNet-50 as the backbone network of CNN. ResNet-50 includes convolutional layers, residual blocks and global average pooling layers. The convolutional layer includes a first convolutional layer Conv1 and a second convolutional layer Conv2. The first convolutional layer Conv1 includes 64 3*3 convolutional kernels and uses the ReLU activation function. The residual blocks, including the second residual block Res2, the third residual block Res3, the fourth residual block Res4, and the fifth residual block Res5, form the core structure of ResNet-50. The global average pooling layer, located at the end of ResNet-50, reduces the multidimensional image feature map to a 1024-dimensional image feature vector. The LSTM network in the environmental parameter encoding module includes a first LSTM layer and a second LSTM layer. Both the first LSTM layer and the second LSTM layer include 256 hidden units for processing time-series information. The fully connected network in the Gamma prediction module includes a first fully connected layer and a second fully connected layer, each containing 256 neurons.
6. The deep learning-based Gamma debugging system according to claim 5, characterized in that: The image feature extraction module uses a convolutional neural network (CNN) to extract a 1024-dimensional image feature vector from the displayed image, including RGB data, image brightness, image color, and image texture. The environmental parameter encoding module uses an LSTM network to encode environmental parameters into a 128-dimensional environmental feature vector. The Gamma prediction module uses a fully connected network to output 256-dimensional optimal Gamma lookup table parameters based on the fusion features.
7. The deep learning-based Gamma debugging system according to claim 4, characterized in that: The training method for the deep learning prediction engine includes: S11. Collect 10,000 high-quality images containing different scenes, obtain the environmental parameters of each image, and obtain the standard Gamma parameters of each image through professional calibration to construct a historical dataset. S12. Divide the historical dataset into training set, validation set and test set according to the preset ratio; S13. Set the loss function and optimizer for the deep learning prediction engine; S14. Input the training set into the deep learning prediction engine to train the model; S15. Calculate the loss value based on the loss function, and the optimizer updates the model parameters based on the loss value and network gradient information; S16. If the loss value is less than the preset threshold, the model training ends and the current deep learning prediction engine is the trained deep learning prediction engine; otherwise, return to S14 and continue to train the model using the training set. S17. Input the validation set into the trained deep learning prediction engine, evaluate the model's generalization ability by observing its performance on the validation set, and fine-tune the model's hyperparameters and structure. S18. Input the test set into the optimized deep learning prediction engine to evaluate the model performance.
8. The deep learning-based Gamma debugging system according to claim 7, characterized in that: The loss function is: Loss function = content loss + perceptual loss + adversarial loss; Among them, content loss is pixel-level error based on PSNR and MSE, perceptual loss is perceptual similarity loss based on VGG network, and adversarial loss is loss from image quality adversarial training based on GAN network.
9. The deep learning-based Gamma debugging system according to claim 1, characterized in that: The hardware execution layer includes a display driver chip, a register configuration module, a temperature compensation module, and a display device; The display driver chip integrates an AI acceleration unit, deploys a deep learning prediction engine and a dynamic compensation module, and supports real-time deep learning inference. Register configuration module, dynamically updates Gamma lookup table parameters; The temperature compensation module compensates for the impact of the display panel's operating temperature on the Gamma lookup table parameters.
10. A deep learning-based Gamma tuning method, applied to the deep learning-based Gamma tuning system of claim 1, characterized in that: Includes the following steps: S1. Data Acquisition: Acquire the current display image of the display device, as well as the corresponding environmental parameters, and analyze the brightness histogram of the display image; S2. Feature Extraction: Extract image features, including RGB data, image brightness, image color, and image texture, from the displayed image, encode environmental parameters into environmental features, and calculate the statistical features of the displayed image. S3, Gamma Prediction: Based on the attention mechanism and feature concatenation operation, image features and environmental features are weighted and fused to obtain fused features. The optimal Gamma lookup table parameters are output based on the fused features. S4. Parameter Verification: Smooth the parameters of the optimal Gamma lookup table, verify their rationality, check whether they exceed hardware limitations, and perform boundary constraint processing. S5. Hardware Update: Dynamically update the parameters of the Gamma lookup table, configure them, and verify the update results; S6. Effect Evaluation: Calculate the adjusted display image quality, record historical adjustment data, and provide feedback for model optimization.