Intelligent Adaptive Projection Lamp and Projection Control Method

Abstract

The present application provides an intelligent adaptive projection lamp and a projection control method. The method includes: collecting multi-source environmental data, where the multi-source environmental data includes at least one of environmental light data, projection surface characteristic data, and user operation data; preprocessing, feature extraction, and feature fusion are performed on the multi-source environmental data to generate an environmental feature vector; through a parameter prediction model based on rate-distortion optimization, the environmental feature vector is projected and mapped to generate projection parameters; image deblurring processing and adaptive k-space sampling reconstruction are performed on the projection parameters to generate an optimized projection image; collecting projection effect feedback data for the optimized projection image; and online updating and optimizing the parameter prediction model according to the projection effect feedback data. The present application realizes the intelligent adaptive adjustment of projection parameters and improves the projection quality.

Description

Technical Field

[0001] This application relates to the technical fields of artificial intelligence and image processing, and particularly relates to an intelligent adaptive projection lamp and a projection control method. Background Art

[0002] With the rapid development of artificial intelligence technology, intelligent projection devices have been widely used in fields such as office work, education, and entertainment. The projection effect of projection devices directly affects the user experience. Therefore, how to achieve intelligent optimization and adjustment of projection parameters has become the focus of attention in the industry.

[0003] Currently, common projection devices mainly adopt projection correction methods based on fixed parameters or simple adaptive adjustment schemes. The projection correction method based on fixed parameters optimizes the projection effect through preset parameters such as brightness and contrast, but it cannot perform adaptive adjustment for different environments. The simple adaptive adjustment scheme only considers a single environmental factor (such as ambient light) for parameter adjustment and cannot achieve multi-dimensional collaborative optimization.

[0004] Some improved schemes introduce machine learning technology to optimize projection parameters through traditional machine learning models. For example, the prior art discloses a method for adaptively adjusting projection parameters based on a support vector machine. This method realizes the prediction and optimization of projection parameters by constructing a support vector machine model. However, due to the limited expression ability of traditional machine learning models, it is difficult to achieve precise control in complex scenarios, resulting in an unsatisfactory projection effect.

[0005] Therefore, how to achieve intelligent and precise adjustment of projection parameters and improve projection quality has become a technical problem to be solved urgently. Summary of the Invention

[0006] In view of this, this application provides an intelligent adaptive projection lamp and a projection control method, which can achieve intelligent and precise adjustment of projection parameters and improve projection quality.

[0007] An embodiment of this application provides an intelligent adaptive projection control method, including:

[0008] Collecting multi-source environmental data, where the multi-source environmental data includes at least one of environmental light data, projection surface characteristic data, and user operation data;

[0009] Preprocessing, feature extraction, and feature fusion are performed on the multi-source environmental data to generate an environmental feature vector;

[0010] The environmental feature vector is projected through a parameter prediction model based on rate distortion optimization to generate projection parameters;

[0011] Perform image deblurring processing and adaptive k-space sampling reconstruction on the projection parameters to generate an optimized projection image;

[0012] Collect projection effect feedback data for the optimized projection image;

[0013] Online update and optimize the parameter prediction model according to the projection effect feedback data.

[0014] Optionally, the acquisition of multi-source environmental data includes:

[0015] Collect the environmental light data through an environmental light sensor, and the environmental light data includes environmental light intensity data;

[0016] Obtain the projection surface characteristic data through a projection surface detector, and the projection surface characteristic data includes the reflectivity and texture feature data of the projection surface;

[0017] Collect the adjustment operation data of the user for the projection parameters through a user interface to obtain the user operation data.

[0018] Optionally, the preprocessing, feature extraction, and feature fusion of the multi-source environmental data to generate an environmental feature vector include:

[0019] Perform noise filtering and outlier detection processing on the environmental light data to obtain the cleaned environmental light data;

[0020] Perform geometric correction and texture feature extraction on the projection surface characteristic data to obtain standardized surface feature data;

[0021] Perform time series alignment and behavior pattern analysis processing on the user operation data to obtain normalized user operation data;

[0022] Use a convolutional neural network to extract the standardized surface feature data to obtain image features;

[0023] Use a recurrent neural network to analyze the normalized user operation data to obtain time series features;

[0024] Through an attention mechanism, fuse the image features, the time series features, and the cleaned environmental light data, and the fused data is the environmental feature vector.

[0025] Optionally, the parameter prediction model is a multi-layer deep neural network, and the multi-layer deep neural network includes a feature receiving layer, multiple fully connected layers based on the rectified linear unit activation function, and a parameter prediction layer;

[0026] The step of performing projection mapping on the environmental feature vector to generate projection parameters by using a parameter prediction model based on rate distortion optimization includes:

[0027] Receiving the environment feature vector through the feature receiving layer;

[0028] In the fully connected layer, a deep representation corresponding to the environment feature vector is generated through a nonlinear transformation based on a ReLU activation function and a residual connection;

[0029] In the parameter prediction layer, projection mapping is performed on the deep representation corresponding to the environmental feature vector based on multiple independent prediction units to generate the projection parameters, wherein one prediction unit corresponds to one projection parameter.

[0030] Optionally, a parameter prediction model based on rate-distortion optimization is obtained in the following manner;

[0031] Construct training data sets and validation sets based on pre-collected historical environmental data and corresponding optimal projection parameters;

[0032] Constructing a loss function based on rate-distortion theory, wherein the loss function includes a parameter prediction error term and a rate constraint term;

[0033] Using an improved stochastic gradient descent algorithm, based on the loss function and the training data set, to train the preset neural network model, and during the training process, using an adaptive moment estimation optimizer and a dynamic learning rate adjustment strategy;

[0034] The model effect is evaluated based on the validation set, and hyperparameters including batch size and learning rate are dynamically adjusted to generate the parameter prediction model.

[0035] Optionally, performing image deblurring processing and adaptive k-space sampling reconstruction on the projection parameters to generate an optimized projection image includes:

[0036] A neural network encoder based on prior knowledge encodes the projection image corresponding to the projection parameters to obtain an implicit representation of the feature encoding of the projection image;

[0037] A decoder network based on a multi-scale structure restores the implicit representation of the feature encoding by layer-by-layer feature reconstruction to obtain a reconstructed image corresponding to the projected image;

[0038] The reconstructed image is fused with the projection image through a residual learning strategy to generate a clear projection image;

[0039] Based on a learnable sampling pattern generator, an adaptive sampling operation is performed according to the content characteristics of the clear projection image to determine the optimal sampling trajectory;

[0040] Use a reconstruction model with a deep convolutional network based on the U-Net architecture to retain image details at different scales through skip connections;

[0041] Based on the Fourier consistency constraint, perform a consistency operation on the optimal sampling trajectory and image details at different scales to generate an initial optimized projection image;

[0042] Adopt an image registration network for precise alignment, and perform geometric correction on the initial optimized projection image to generate the optimized projection image.

[0043] Optionally, in the step of obtaining the reconstructed image corresponding to the projection image, the projection control method further includes: adopting a spatial attention mechanism to enhance the detail performance of important regions, and adding an adaptive weight adjustment strategy to dynamically adjust the deblurring intensity according to the blur degree of the image region.

[0044] Optionally, the collecting of the projection effect feedback data for the optimized projection image includes:

[0045] Collect the brightness uniformity, color restoration degree, and contrast of the projection image through an optical sensor to form a time-series data stream;

[0046] Determine the quantitative indicators of the projection image through a preset evaluation algorithm, where the quantitative indicators include peak signal-to-noise ratio and structural similarity;

[0047] Collect user evaluation data through a user interface, where the user evaluation data includes a satisfaction score and feedback information;

[0048] Record the operation behavior data of the user's parameter adjustment;

[0049] Perform data cleaning and standardization processing on the time-series data stream, the quantitative indicators, the user evaluation data, and the operation behavior data to obtain standardized projection effect feedback data.

[0050] Optionally, the online update and optimization of the parameter prediction model according to the projection effect feedback data includes:

[0051] Based on the standardized projection effect feedback data, adopt an incremental learning strategy, and gradually update the parameters of the parameter prediction model through the mini-batch gradient descent method. And during the parameter update process, based on the dynamic learning rate adjustment mechanism, adjust the parameter prediction model according to the consistency degree of the standardized projection effect feedback data, and based on the knowledge distillation technology, transfer the empirical knowledge of the historical model to the updated parameter prediction model;

[0052] By comparing the projection quality of the projection images before and after updating the parameter prediction model, and based on the comparison result, verify the degree of performance improvement of the parameter prediction model, and decide whether to accept this update based on the comparison result.

[0053] An embodiment of the present application further provides an intelligent adaptive projection lamp, including a processor and a memory. The memory is used to store a computer program. When the processor executes the computer program, the above-mentioned projection control method is implemented.

[0054] An embodiment of the present application further provides a computer device, which includes:

[0055] At least one processor; and,

[0056] A memory communicatively connected to the at least one processor; wherein,

[0057] The memory stores instructions executable by the at least one processor. When the instructions are executed by the at least one processor, the at least one processor can execute the above-mentioned intelligent adaptive projection control method.

[0058] An embodiment of the present application further provides a computer-readable storage medium, which stores computer instructions for causing a computer to execute the above-mentioned intelligent adaptive projection control method.

[0059] An embodiment of the present application further provides a computer program product, including computer instructions, which implement the steps of the above-mentioned intelligent adaptive projection control method when executed by a processor.

[0060] The present application has the following technical effects:

[0061] By adopting a rate-distortion optimized parameter prediction model, combining the deep feature extraction and fusion of multi-source environmental data, accurate prediction and adaptive adjustment of projection parameters are achieved, significantly improving the projection quality. The implicit neural representation method based on prior knowledge and the deep end-to-end K-space processing scheme effectively improve the image deblurring and reconstruction effects. Through the online learning and model optimization mechanism, the system can continuously learn and evolve, and continuously improve the projection effect. In other words, the present application can achieve intelligent and accurate adjustment of projection parameters and improve the projection quality. Description of the Drawings

[0062] In order to more clearly illustrate the disclosed embodiments in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative efforts.

[0063] Figure 1 It is a schematic flowchart of the intelligent adaptive projection control method provided by the embodiment of the present application;

[0064] Figure 2 It is a structural diagram of the intelligent adaptive projection lamp in the embodiment of the present application. Specific implementation manners

[0065] The following describes the embodiments of the present disclosure in detail with reference to the accompanying drawings.

[0066] It should be clear that the following uses specific specific examples to illustrate the implementation manners of the present disclosure. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without making creative efforts belong to the scope of protection of the present disclosure.

[0067] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present disclosure, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, this device and / or this method can be implemented using other structures and / or functions in addition to one or more of the aspects described herein.

[0068] It also should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present disclosure in a schematic manner. The drawings only show the components related to the present disclosure, rather than being drawn according to the number, shape and size of the components in actual implementation. The type, quantity and proportion of each component in its actual implementation can be an arbitrary change, and the component layout type may also be more complex.

[0069] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0070] The following will explain in detail the specific implementation manners of the present application in conjunction with the accompanying drawings. It should be noted that the accompanying drawings of the specification of the present application are only used for the interpretation and illustration of the present application, and are not intended to limit the protection scope of the present application.

[0071] The technical solutions of the embodiments of the present application will be described in detail in conjunction with the accompanying drawings.

[0072] It should be noted that the technical terms used in the embodiments of the present application should be understood as having the meanings commonly understood by those skilled in the art, unless otherwise clearly defined and limited. The descriptions such as "first", "second", etc. used in the embodiments of the present application are only used to distinguish different entities or operations, and should not be construed as limiting the specific order of these entities or operations.

[0073] The following in conjunction with Figure 1 will describe in detail the specific implementation manners of the embodiments of the present application.

[0074] As Figure 1 shown, the embodiments of the present application provide an intelligent adaptive projection control method. The method includes steps such as multi-source environmental data acquisition, feature extraction and fusion, parameter prediction, image optimization processing, and online learning and update.

[0075] It should be noted that the present application uses deep learning technology to achieve intelligent optimization of projection parameters, which can effectively improve the projection quality.

[0076] More specifically, the intelligent adaptive projection control method includes:

[0077] S1: Collect multi-source environmental data, where the multi-source environmental data includes at least one of environmental light data, projection surface characteristic data, and user operation data.

[0078] Specifically, in step S1, first perform the step of collecting multi-source environmental data, and this step comprehensively obtains environmental information through multiple data collection modules.

[0079] Exemplarily, an environmental light sensor collects environmental light intensity data to monitor the light change of the projection environment in real time.

[0080] A projection surface detector obtains the reflectivity and texture feature data of the projection surface, providing an important basis for subsequent parameter optimization. In addition, the adjustment operation data of the user is collected through the user interaction interface to understand the personalized needs of the user.

[0081] Step S1 specifically includes:

[0082] S1.1: Through the environmental light sensor, collect the environmental light data, where the environmental light data includes environmental light intensity data;

[0083] In step S1.1, ambient light data is collected by an ambient light sensor.

[0084] It should be noted that this sensor uses a high-precision photosensitive element array and can capture the light changes in the environment in real time.

[0085] Specifically, the ambient light sensor can measure the light intensity of the projection area and its surrounding environment at a sampling frequency of 10 times per second. Among them, the data collected by the ambient light sensor includes information such as the intensity value of the ambient light (in lux), the light distribution uniformity, and the light fluctuation trend.

[0086] In addition, the ambient light sensor in this application also has an anti-glare processing function, which can effectively filter the interference of strong light sources and ensure the accuracy of the collected data.

[0087] S1.2: Obtain the projection surface characteristic data through a projection surface detector, where the projection surface characteristic data includes the projection surface.

[0088] In step S1.2, a projection surface detector is used to obtain the projection surface characteristic data. First, the projection surface detector uses a high-resolution image sensor to perform multi-angle scanning and collection of the projection surface. Second, through the built-in spectral analysis module, the reflectivity data of the projection surface is measured, including the diffuse reflectivity and the specular reflectivity.

[0089] It should be particularly noted that when obtaining the projection surface characteristic data, the texture feature data of the projection surface is also collected. The texture feature data can include information such as the surface roughness, color uniformity, and material change. The collection of these characteristic data provides an important basis for subsequent parameter optimization.

[0090] S1.3: Collect the adjustment operation data of the user for the projection parameters through a user interface to obtain the user operation data.

[0091] The user interface provides intuitive parameter adjustment controls, including a brightness slider, a contrast knob, a color temperature selector, etc. When the user adjusts these controls, information such as the parameter type, adjustment direction, adjustment amplitude, and adjustment time of the adjustment will be recorded.

[0092] In addition, the user interface is also provided with quick operation buttons to facilitate the user to quickly switch the preset parameter combinations.

[0093] It should be noted that this solution not only records the specific operation data of the user, but also analyzes the user's operation habits and preferences to provide a reference for personalized parameter optimization.

[0094] It should be noted that in the actual operation process, the projection parameters can be determined based on at least two of the environmental light data, the projection surface characteristic data, and the user operation data.

[0095] S2: Preprocess, extract features, and fuse features from the multi-source environmental data to generate an environmental feature vector.

[0096] It should be noted that multiple deep learning models are used for feature extraction and fusion in this step.

[0097] First, filter the noise of the environmental light data to eliminate the influence of outliers. Second, perform geometric correction on the projection surface characteristic data to ensure the accuracy of feature extraction.

[0098] Specifically, a convolutional neural network is used to extract image features. This network contains multiple convolutional layers and pooling layers, which can effectively capture the spatial features of the surface texture. At the same time, a recurrent neural network is used to analyze the temporal features of user operations, and the user behavior pattern is modeled through LSTM units. Finally, an attention mechanism is introduced to achieve feature fusion, which can adaptively adjust the weights of different features to generate a high-quality environmental feature vector.

[0099] S2 specifically includes:

[0100] S2.1: Filter the noise and detect outliers of the environmental light data to obtain the cleaned environmental light data.

[0101] In step S2.1, filter the noise and detect outliers of the collected environmental light data.

[0102] First, use a Gaussian filter to preprocess the environmental light data to eliminate the influence of random noise. Second, perform outlier detection based on statistical methods, calculate the mean and standard deviation of the light data, and mark the data points that deviate from the mean by more than a preset multiple of the standard deviation (for example, three times the standard deviation) as potential outliers.

[0103] In addition, considering the dynamic change characteristics of environmental light, the system also introduces an adaptive threshold mechanism, which can dynamically adjust the outlier determination standard according to the light change law at different times. Through these processes, high-quality cleaned environmental light data is finally obtained.

[0104] It should be noted that these outliers are not directly deleted in this solution, but are corrected using an interpolation method based on temporal context to expand the quantity of environmental light data.

[0105] S2.2: Perform geometric correction and texture feature extraction on the projection surface characteristic data to obtain standardized surface feature data.

[0106] Specifically, first, a perspective transformation algorithm is used to correct the geometric distortion of the projection surface to ensure the accuracy of feature extraction. Second, a multi-scale Gabor filter bank is used to extract the texture features of the projection surface. This filter bank contains filter kernels with different directions and scales, which can comprehensively capture the texture information of the surface.

[0107] It should be particularly noted that during the texture feature extraction process, the non-uniformity of the surface material is also considered. Through the adaptive local feature enhancement technology, the robustness of feature extraction is improved. After these processes, the system obtains standardized surface feature data, laying a foundation for subsequent feature fusion.

[0108] S2.3: Perform temporal alignment and behavioral pattern analysis on the user operation data to obtain normalized user operation data.

[0109] Optionally, in S2.3, the dynamic time warping algorithm is used to align the operation sequences at different time scales to ensure the consistency of temporal features. Second, the user operations are segmented through the sliding window technique, so that each window contains 30 seconds of continuous operation data.

[0110] It should be noted that a hierarchical feature extraction strategy is adopted in the behavioral pattern analysis, including low-level operation frequency statistics, middle-level operation combination pattern recognition, and high-level user preference inference. Through this multi-level analysis and processing, normalized user operation data is finally obtained.

[0111] S2.4: Use a convolutional neural network to extract the standardized surface feature data to obtain image features.

[0112] Among them, this convolutional neural network adopts an improved ResNet structure, which contains five convolutional blocks. Each convolutional block consists of two 3×3 convolutional layers and a max-pooling layer.

[0113] In addition, considering the diversity of the projection surface features, an adaptive feature normalization layer is also added to the convolutional neural network. The adaptive feature normalization layer can dynamically adjust the normalization parameters according to the feature distributions of different surfaces. Thus, through this deep network structure, rich image features can be successfully extracted.

[0114] It should be particularly noted that in order to improve the effect of feature extraction, an attention mechanism and a residual connection can be introduced into each convolutional block.

[0115] S2.5: Use a recurrent neural network to analyze the normalized user operation data to obtain temporal features.

[0116] First, a bidirectional LSTM network structure is adopted. This network contains two layers of LSTM units, and each layer contains 128 hidden states. Secondly, in order to improve the memory ability of the network, the system also adopts a residual time series connection structure, enabling the network to better capture long-term dependence relationships. Through the analysis of this recurrent neural network, the system obtains the time series feature representation of the user's operation behavior.

[0117] It should be noted that a gated attention mechanism is introduced into the LSTM unit, and the gated attention mechanism can adaptively focus on important time series features.

[0118] S2.6: Through the attention mechanism, fuse the image features, the time series features, and the cleaned environmental light data. The fused data is the environmental feature vector.

[0119] In step S2.6, the image features, the time series features, and the cleaned environmental light data are fused through the attention mechanism.

[0120] Specifically, first, a multi-head self-attention mechanism is used to adaptively weight the features of different modalities. Secondly, the features of different dimensions are mapped to the same feature space through a feature transformation network. And, in order to improve the expression ability of the fused features, a residual feature fusion structure is also adopted to combine the original feature information with the transformed features. Finally, through this multi-modal feature fusion strategy, the system generates high-quality environmental feature vectors, providing reliable inputs for subsequent parameter prediction.

[0121] It should be particularly noted that a dynamic weight adjustment mechanism is introduced in the feature fusion process, and this mechanism can automatically adjust the importance weights of different features according to the characteristics of the current scene.

[0122] S3: Through a parameter prediction model based on rate-distortion optimization, projectively map the environmental feature vector to generate projection parameters.

[0123] In step S3, a parameter prediction model based on rate-distortion optimization is used for projective mapping.

[0124] Exemplarily, this parameter prediction model adopts a multi-layer deep neural network structure, and the multi-layer deep neural network includes a feature receiving layer, multiple fully connected layers based on the rectified linear unit activation function, and a parameter prediction layer.

[0125] Specifically, first, the feature receiving layer receives the environmental feature vector to maintain the integrity of the input data. Secondly, the ReLU activation function is used in the fully connected layer for non-linear transformation, and at the same time, residual connections are introduced to alleviate the problem of gradient disappearance.

[0126] It should be noted that the parameter prediction layer contains multiple independent prediction units, and each unit is responsible for predicting a specific projection parameter (such as brightness, contrast, etc.). In addition, the model training uses a loss function based on the rate distortion theory, which comprehensively considers the accuracy of parameter prediction and computational complexity, and can optimize the use of computing resources while ensuring the prediction effect.

[0127] For ease of understanding, a specific example is used to illustrate the working process of the parameter prediction model.

[0128] Suppose in a conference room environment, the ambient light intensity is 800 lux and the reflectivity of the projection surface is 0.75. First, encode these ambient light intensity and projection surface reflectivity into a feature vector. Then, the feature vector passes through the first hidden layer with 64 neurons and undergoes a non-linear transformation through the ReLU activation function. Next, the information is passed to the second hidden layer containing 128 neurons to further extract deep features. Finally, multiple prediction units in the output layer generate different projection parameters respectively, such as adjusting the brightness to 75% and setting the contrast to 1.8, etc. This multi-level feature extraction and mapping process ensures the accurate prediction of projection parameters.

[0129] S3 specifically includes:

[0130] S3.1: Receive the ambient feature vector through the feature receiving layer.

[0131] In step S3.1, receive the ambient feature vector through the feature receiving layer.

[0132] It should be noted that the feature receiving layer adopts an innovative adaptive feature normalization structure, which can dynamically adjust the normalization parameters according to the statistical characteristics of the input features.

[0133] Specifically, first perform dimension checking and alignment on the input ambient feature vector to ensure that the feature dimensions match the network structure. Secondly, through batch normalization processing, standardize the feature distribution to a suitable range, which is crucial for improving the stability of subsequent processing. In addition, the feature receiving layer also integrates a feature integrity verification mechanism. When detecting missing or abnormal features, it will trigger the feature completion module to reasonably estimate the missing features through context information.

[0134] For example, when there is a temporary missing of ambient light features, interpolation will be performed based on historical data and feature values at adjacent moments for supplementation.

[0135] S3.2: In the fully connected layer, generate a deep representation corresponding to the ambient feature vector through a non-linear transformation based on the ReLU activation function and residual connection.

[0136] It should be noted that the fully connected layer adopts a multi-layer structure design, including three hidden layers with the number of neurons being 256, 512, and 256 respectively. Through this progressive structure design, the abstract representation of features can be gradually extracted.

[0137] It should be particularly noted that a ReLU activation function is connected after each hidden layer. This activation function can not only introduce the ability of nonlinear transformation but also effectively alleviate the problem of gradient disappearance. Secondly, this solution also innovatively adds residual connections between adjacent layers. These skip connections can ensure the lossless transmission of information and also provide more optimization paths for the network.

[0138] In addition, in order to further improve the effect of feature extraction, this solution also introduces a Dropout mechanism after each fully connected layer. By randomly deactivating some neurons, overfitting of the model is effectively prevented.

[0139] For example, during the training process, the Dropout ratio is set to 0.3, that is, 30% of the neurons are randomly deactivated during each forward propagation.

[0140] S3.3: In the parameter prediction layer, based on multiple independent prediction units, a projection mapping is performed on the deep representation corresponding to the environmental feature vector to generate the projection parameter, where one prediction unit corresponds to one projection parameter.

[0141] Specifically, the parameter prediction layer contains multiple parallel prediction branches, and each branch is responsible for the prediction of a specific projection parameter.

[0142] It should be noted that each prediction unit adopts a dedicated network structure and is optimized according to the characteristics of the prediction parameter. For example, for the prediction unit of the brightness parameter, a three-layer neural network is used, and the Sigmoid activation function is used in the output layer to ensure that the predicted value falls within a reasonable range; for the prediction unit of the contrast parameter, a two-layer structure with residual connections is used, and the range mapping is achieved through the Tanh activation function.

[0143] In addition, this solution also introduces a parameter constraint mechanism in the prediction layer. By adding a regularization term, the coordination between prediction parameters is ensured. To improve the robustness of the prediction, the system also designs a prediction confidence evaluation module. When the prediction confidence is lower than the threshold, a backup prediction strategy will be triggered, and a more reliable output will be provided by integrating multiple prediction results.

[0144] For example, in practical applications, if the prediction confidence of a certain parameter is lower than 0.8, multiple prediction models will also be adopted simultaneously, and the final predicted value will be obtained through weighted averaging.

[0145] For ease of understanding, take a specific scenario in a projection environment as an example.

[0146] Suppose in a conference room environment, the environmental feature vector contains the following information: {ambient light intensity: 800 lux, projection surface reflectivity: 0.75, projection distance: 3 meters, room temperature: 25 °C, user's preferred brightness level: 4}. The prediction model adopts a five-layer neural network structure, and the design of each layer has its specific purpose.

[0147] First, the input layer contains 32 neurons for receiving the above environmental feature vector. Secondly, the three hidden layers contain 64, 128, and 64 neurons respectively. Each layer uses the ReLU activation function and introduces residual connections. Taking the first hidden layer as an example, it can convert raw features such as ambient light intensity and surface reflectivity into higher-level representations, such as "ambient brightness suitability index". Finally, the output layer is designed with 4 prediction units, corresponding to the brightness (0 - 100%) of the projection, contrast (1.0 - 2.0), color temperature (2000K - 6500K), and sharpness (0 - 10) respectively. Through this structure, the model can map the input environmental feature vector to the optimal combination of projection parameters, for example: {brightness: 75%, contrast: 1.8, color temperature: 5500K, sharpness: 8}.

[0148] Taking the above conference room scenario as an example, first construct the loss function L = α * L_pred + β * L_rate + γ * (L1 + L2), where L_pred represents the parameter prediction error, L_rate represents the rate constraint, L1 and L2 are regularization terms, and α, β, γ are weight coefficients. Specifically, when the model predicts the projection brightness to be 75% while the actual optimal brightness is 72%, the L_pred term will calculate this 3% deviation. At the same time, the L_rate term will evaluate the computational complexity of the model when giving this prediction. For example, if the model uses too many parameters or layers, the L_rate value will increase. In addition, a dynamic batch strategy is adopted during training, with the initial batch size being 32 and the initial learning rate set to 0.001. It should be particularly noted that to improve the robustness of the model, different interference situations are added to the training data: for example, by randomly adjusting the ambient light intensity by ±100 lux or changing the projection surface reflectivity by ±0.1 to simulate different usage scenarios. Finally, the model parameters are updated through the backpropagation algorithm, and the training stops when the mean squared error on the validation set is less than 0.05.

[0149] S4: Perform image deblurring processing and adaptive k-space sampling reconstruction on the projection parameters to generate an optimized projection image.

[0150] This step uses advanced image processing techniques to improve the projection quality.

[0151] Specifically, first, the projection image is encoded into an implicit representation by a neural network encoder based on prior knowledge, which can effectively capture the essential features of the image. Then, a decoder network with a multi-scale structure is used for image reconstruction, and a spatial attention mechanism is introduced during the reconstruction process to enhance the detail performance of important regions. In addition, a learnable sampling pattern generator is adopted to determine the optimal sampling trajectory, and high-quality image reconstruction is achieved through a deep convolutional network based on the U-Net architecture.

[0152] Exemplarily, take processing a commercial presentation slide with a resolution of 1920×1080 as an example.

[0153] First, a prior knowledge base containing common font and graphic features is constructed. The encoder encodes the input image into a 256-dimensional feature vector, paying particular attention to the feature expression in the blurred areas. For the text area with a font size of 24pt, a higher weight (e.g., weight value 0.8) is assigned through the spatial attention mechanism, while a lower weight (e.g., weight value 0.2) is assigned to the solid-color background area. The decoder gradually restores the image details through 3 upsampling blocks, and each block contains a residual learning unit. Finally, through an adaptive weight fusion module, the de-blurred image and the original image are fused in a ratio of 0.7:0.3, significantly improving the image clarity.

[0154] S4 specifically includes:

[0155] S4.1: A neural network encoder based on prior knowledge encodes the projection image corresponding to the projection parameters to obtain an implicit representation of the feature encoding of the projection image.

[0156] This encoder adopts an innovative knowledge-driven architecture design, which can effectively utilize the prior knowledge of image features accumulated in advance.

[0157] Specifically, first, a knowledge base containing common image feature templates is constructed, and these templates cover the standard feature representations of various visual elements such as text, graphics, and charts. Secondly, the encoder extracts features through a multi-layer convolutional network structure, and each layer is equipped with an attention mechanism, which can adaptively adjust the focus of feature extraction according to the content characteristics. In addition, a feature consistency constraint is introduced during the encoding process to ensure that the encoding result is highly correlated with the prior knowledge.

[0158] For example, when processing an area containing text, the encoder will pay particular attention to the stroke features and edge information of the font, and generate more accurate feature encoding by matching with the standard font features in the prior knowledge base.

[0159] S4.2: Based on the decoder network with a multi-scale structure, through layer-by-layer feature reconstruction, the implicit representation of the feature encoding is restored to obtain the reconstructed image corresponding to the projected image.

[0160] In step S4.2, a decoder network with a multi-scale structure is used for feature reconstruction.

[0161] First of all, the decoder adopts a hierarchical reconstruction strategy, including four reconstruction stages, and each stage is equipped with feature reconstruction modules of different scales.

[0162] It should be noted that the system introduces a spatial attention mechanism in each reconstruction stage, which can dynamically adjust the reconstruction intensity according to the importance of the image content.

[0163] In addition, in order to improve the reconstruction effect, the system also designs an adaptive weight adjustment strategy, which can automatically adjust the reconstruction parameters according to the blur degree of the image area. For example, for high-frequency detail areas such as text areas, a higher reconstruction weight will be given to ensure the clear restoration of details.

[0164] S4.3: Through the residual learning strategy, the reconstructed image and the projected image are fused to generate a clear projected image.

[0165] Specifically, a multi-level residual connection network is first established, which can capture the difference information between the original image and the reconstructed image at different scales.

[0166] It should be noted that the system uses an adaptive weighting mechanism for image fusion, and the fusion weights will be dynamically adjusted according to the content features and quality indicators of the image area.

[0167] In addition, the system also introduces a local enhancement processing module, which can perform refined fusion optimization for specific areas. For example, when processing high-contrast edge areas, the system will appropriately increase the weight of the original image to retain more edge details.

[0168] S4.4: Based on the learnable sampling pattern generator, according to the content features of the clear projected image, perform an adaptive sampling operation to determine the optimal sampling trajectory.

[0169] First of all, the generator adopts a deep reinforcement learning framework to optimize the sampling strategy through continuous interaction with the environment.

[0170] It should be noted that the system designs a content-aware reward function, which can dynamically adjust the sampling parameters according to the quality features of the sampling results.

[0171] Secondly, the system also introduces a sampling efficiency evaluation mechanism to ensure the efficiency of the sampling process by monitoring the sampling quality and computational overhead in real time. For example, for areas with complex textures, the system will automatically increase the sampling density, while for flat areas, it will appropriately reduce the sampling frequency.

[0172] S4.5: Use a reconstruction model based on a deep convolutional network with a U-Net architecture to retain image details at different scales through skip connections.

[0173] Specifically, the network adopts a symmetric encoder-decoder structure, including five downsampling and five upsampling modules.

[0174] It should be noted that the system sets skip connections between each layer, and these connections can effectively retain the multi-scale feature information of the image.

[0175] In addition, the system also adds a feature selection mechanism to the network, which can automatically identify and retain key image details. For example, during the reconstruction process, the system will pay special attention to the retention of high-frequency information such as edges and textures to ensure the clarity of the reconstructed image.

[0176] S4.6: Based on the Fourier consistency constraint, perform a consistency operation on the optimal sampling trajectory and image details at different scales to generate an initial optimized projection image.

[0177] First, the system transforms the sampling data into the frequency domain space and analyzes the frequency characteristics of the image through Fourier transform.

[0178] It should be particularly noted that the system designs an adaptive frequency filter, which can dynamically adjust the retention ratio of frequency components according to the content characteristics of the image.

[0179] Secondly, the system also introduces a phase consistency constraint to ensure that the reconstruction results are highly consistent in both the frequency domain and the spatial domain. For example, the system will pay special attention to the smooth transition of low-frequency and medium-frequency components while retaining sufficient high-frequency details to achieve the best visual effect.

[0180] S4.7: Use an image registration network for precise alignment, perform geometric correction on the initial optimized projection image to generate the optimized projection image.

[0181] The network adopts an end-to-end learning architecture and can automatically estimate and correct the geometric transformation parameters of the image.

[0182] It should be noted that the system introduces a multi-scale feature matching strategy during the registration process to improve the accuracy of registration by performing feature matching at different resolution levels.

[0183] In addition, the system also designs a registration quality assessment module, which can monitor the registration effect in real time and trigger optimization and adjustment when necessary. For example, when the registration error in a local area is detected to exceed the threshold, the system will automatically start the local fine registration program to ensure the consistency of the overall registration effect.

[0184] Take the projection of a commercial presentation slide with a resolution of 1920×1080 as an example.

[0185] It should be noted that the slide contains various elements such as text, charts, and logos. Due to the influence of the projection distance and angle, the original image appears blurred in some areas.

[0186] First, construct a prior knowledge base, which contains clear feature templates of common fonts and graphics, such as the stroke features of Arial font and the edge features of common charts. Based on these prior knowledge, design an encoder-decoder network structure. Among them, the encoder encodes the input image into a 256-dimensional feature vector, paying special attention to the feature expression of the blurred area. For example, for the text area with a font size of 24pt in the slide, a higher weight (such as a weight value of 0.8) is given through the spatial attention mechanism, while a lower weight (such as a weight value of 0.2) is given to the solid color background area. Then, the decoder network gradually restores the image details through 3 upsampling blocks, and each block contains a residual learning unit. Taking the text area as an example, the blurred font edge is restored to a clear stroke by matching with the prior font features. Finally, through the adaptive weight fusion module, the deblurred image and the original image are fused in a ratio of 0.7:0.3 to obtain an output image with a 30% improvement in clarity.

[0187] First, optimize the K-space sampling. Specifically, the sampling pattern generator will adaptively generate the sampling trajectory according to the image content features. For example, for the area containing important text information (such as the title), a dense sampling strategy is adopted with a sampling rate of 90%; while for non-critical areas such as the background, sparse sampling is adopted and the sampling rate is reduced to 30%, so as to achieve efficient allocation of computing resources.

[0188] It should be particularly noted that in the reconstruction stage, an improved U-Net structure is adopted, which includes 4 layers of downsampling and 4 layers of upsampling, and the number of feature channels increases from 64 to 512.

[0189] Taking a chart in a slide as an example, first, multi-scale features are extracted through downsampling to capture information from detailed textures to the overall structure; then, high-frequency details of the original image are retained through skip connections to ensure the sharpness of the reconstructed chart's edges. In the registration stage, the spatial transformation parameters learned by the registration network include: translation amounts (Δx = 5 pixels, Δy = 3 pixels), rotation angle (θ = 2°), and scaling factor (s = 1.02). Through precise transformation of these parameters, it is ensured that the projected image is perfectly aligned with the target position. In addition, the quality assessment module monitors the peak signal-to-noise ratio (PSNR) of the output image in real time. When the PSNR is lower than 35 dB, a parameter fine-tuning mechanism is triggered.

[0190] S5: Collect projection effect feedback data for the optimized projected image.

[0191] S5 specifically includes:

[0192] S5.1: Through an optical sensor, collect the brightness uniformity, color reproducibility, and contrast of the projected image to form a time-series data stream.

[0193] It should be noted that this sensor uses a high-precision CMOS imaging element and is equipped with a professional-grade optical filter group, which can accurately capture the characteristic parameters of the projected image.

[0194] Specifically, sampling is performed every 100 milliseconds, and key indicators such as brightness uniformity, color reproducibility, and contrast are recorded simultaneously.

[0195] In terms of brightness uniformity measurement, the sensor scans the projection area through a 9-point sampling method to calculate the brightness difference between the central area and the edge area. For the measurement of color reproducibility, a standard color card comparison method is used, and the color accuracy is evaluated by calculating the color difference △E.

[0196] In addition, the system also monitors the contrast change in real time. By measuring the brightness ratio between the brightest area and the darkest area, the dynamic range performance of the image is ensured. These data are organized in chronological order to form a continuous time-series data stream, providing basic support for subsequent quality assessment.

[0197] S5.2: Through a preset evaluation algorithm, determine the quantitative indicators of the projected image, and the quantitative indicators include peak signal-to-noise ratio and structural similarity;

[0198] This solution uses an improved peak signal-to-noise ratio (PSNR) calculation method, which considers the weighted factors of human visual characteristics and can more accurately reflect the image quality.

[0199] It should be particularly noted that in the PSNR calculation process, an adaptive region division strategy is also introduced, and different evaluation weights are used for regions of different content types.

[0200] Secondly, the system evaluates the structural fidelity of the image through the Structural Similarity (SSIM) index, which comprehensively evaluates the image quality from three dimensions: brightness, contrast, and structure. In addition, the system also innovatively designs a temporal consistency evaluation module to timely detect and respond to quality fluctuations by analyzing the quality change trend between consecutive frames.

[0201] For example, when it is detected that the PSNR value drops by more than 3dB within a short period, the system will automatically trigger the quality alarm mechanism.

[0202] S5.3: Collect user evaluation data through the user interface, where the evaluation data includes satisfaction scores and feedback information.

[0203] Specifically, the interface adopts an intuitive scoring system, including a satisfaction scale from 1 to 5 and a detailed feedback input area.

[0204] It should be noted that the system designs an intelligent feedback collection mechanism that not only records the user's direct scores but also captures the user's operation behavior characteristics. For example, when the user frequently adjusts a certain parameter, the system will automatically pop up a relevant feedback inquiry window to understand the user's adjustment intention and expected effect.

[0205] In addition, the system also establishes a semantic analysis module for user feedback, which can extract key information from the user's text description and convert it into quantifiable evaluation indicators.

[0206] S5.4: Record the operation behavior data of the user's parameter adjustment.

[0207] This solution designs a fine-grained operation recording mechanism that can capture each parameter adjustment action of the user, including information such as the type of parameter adjusted, the adjustment direction, the adjustment amplitude, and the adjustment time.

[0208] It should be particularly noted that during the actual processing process, details such as pauses and rollbacks during the user's adjustment will also be recorded, and this information can reflect the user's decision-making process and preference characteristics.

[0209] In addition, an operation mode recognition module is also introduced to identify the commonly used parameter combinations and adjustment habits of the user by analyzing the user's adjustment sequence. For example, the system can discover the parameter setting modes preferred by the user under different lighting environments.

[0210] S5.5: Perform data cleaning and standardization processing on the temporal data stream, the quantitative indicators, the user evaluation data, and the operation behavior data to obtain normalized projection effect feedback data.

[0211] First, an outlier detection algorithm is used to preliminarily screen the data and eliminate the obviously abnormal data points. Secondly, the system performs multi-dimensional standardization processing to convert different types of feedback data into a unified evaluation framework. In addition, the system also establishes a data quality evaluation mechanism to ensure the high credibility of the normalized feedback data by calculating indicators such as data integrity, consistency, and timeliness. For example, for user rating data, the system will combine factors such as the time interval of the rating and the operation context to judge the validity of the rating and adjust its weight in the overall evaluation accordingly.

[0212] It should be noted that the system uses a hybrid method based on statistics and machine learning for data cleaning, considering both the statistical characteristics of the data and the temporal correlation of the data.

[0213] S6: Online update and optimization of the parameter prediction model according to the projection effect feedback data.

[0214] In steps S5 and S6, the system establishes a complete feedback optimization mechanism. This mechanism realizes the continuous improvement of model performance through multi-dimensional data collection and online learning update.

[0215] Exemplarily, the system collects the quality data of the projection image every 100 ms, including parameters such as brightness uniformity and color reproduction. At the same time, satisfaction ratings and feedback information are collected through the user interface. Based on these feedback data, the system adopts an incremental learning strategy to gradually optimize the prediction model and ensures the stability of the update through a dynamic learning rate adjustment mechanism. In addition, knowledge distillation technology is introduced to transfer the empirical knowledge of the historical model to the new model to further improve the model performance.

[0216] S6 specifically includes:

[0217] S6.1: Based on the normalized feedback data, adopt an incremental learning strategy, and through the mini-batch gradient descent method, gradually update the parameters of the parameter prediction model. And during the parameter update process, based on the dynamic learning rate adjustment mechanism, adjust the parameter prediction model according to the consistency degree of the feedback data, and based on knowledge distillation technology, transfer the empirical knowledge of the historical model to the updated parameter prediction model.

[0218] First, the system uses the mini-batch gradient descent method to update the model parameters.

[0219] Specifically, each time 32 samples are selected to form a mini-batch, and the gradient of the loss function is calculated to guide the adjustment direction of the parameters.

[0220] It should be noted that the system adopts a dynamic learning rate adjustment mechanism during the update process, which can adaptively adjust the learning rate according to the consistency degree of the feedback data. For example, when the feedback data of consecutive batches shows a high degree of consistency, the system will appropriately increase the learning rate to accelerate the convergence speed; conversely, when the feedback data shows large fluctuations, the system will decrease the learning rate to ensure the stability of the update.

[0221] In addition, the system innovatively introduces the knowledge distillation technology. By designing a dual architecture of the teacher model and the student model, the effective inheritance of historical knowledge is realized. Specifically, the existing model is used as the teacher model, and the newly trained model is used as the student model. The degree of knowledge transfer is controlled by the temperature parameter T to ensure that the updated model can not only maintain the original generalization ability but also adapt to the new scenario requirements.

[0222] During the incremental learning process, the system also designs a feature importance evaluation mechanism. Specifically, by calculating the contribution degree of different features to the prediction result, the system can identify the key features and focus on optimizing the model parameters related to these features during the update process. It should be particularly noted that the system adopts a feature importance analysis method based on SHAP values, which can accurately quantify the influence of each feature. For example, when the system finds that the environmental light feature has a significant impact on the prediction result, it will correspondingly increase the weight of this feature in the model update. In addition, the system also establishes a verification mechanism for parameter update, and evaluates the effectiveness of the update through cross-validation to ensure the steady improvement of the model performance.

[0223] S6.2: Verify the degree of performance improvement of the parameter prediction model by comparing the projection quality of the projection images before and after the update of the parameter prediction model, and decide whether to accept this update based on the comparison result.

[0224] In step S6.2, the system verifies the effect of the model update through a comprehensive performance evaluation mechanism. First, the system establishes a multi-dimensional evaluation index system, including key indicators such as prediction accuracy, response time, and resource consumption. It should be noted that the evaluation process adopts a sliding window mechanism, and the change trend of the model performance is accurately reflected by comparing the performance indicators within consecutive time windows. Specifically, the system will calculate the change rate of the prediction error before and after the update, the improvement degree of the calculation delay, and the optimization of the resource utilization rate. In addition, the system also designs an adaptive performance threshold mechanism to dynamically adjust the acceptance criteria according to the requirements of different application scenarios. For example, in scenarios with high real-time requirements, the system will pay more attention to the improvement of the response time.

[0225] Based on performance verification, the system has implemented an intelligent update decision-making mechanism. Specifically, the system comprehensively analyzes various performance indicators and calculates the update benefit score. It should be noted that this benefit score is calculated by weighted summation, and the weight coefficients are dynamically adjusted according to the characteristics of the application scenario. For example, in a scenario with high requirements for image quality, the prediction accuracy indicator will be given a higher weight. When the benefit score exceeds the preset threshold, the system will accept this update; otherwise, it will trigger a rollback mechanism to restore to the model state before the update. In addition, the system also maintains a historical record of model versions, supports version switching when necessary, and ensures that the system always runs on the optimal model version.

[0226] The embodiment of the present application also provides an intelligent adaptive projection lamp. As Figure 2 shown, the intelligent adaptive projection lamp 100 may include a processor 110 and a memory 120. The memory 120 is used to store a computer program. When the processor 110 executes the computer program, the above-mentioned projection control method is implemented.

[0227] It should be noted that the projection control method of the embodiment of the present application can also be implemented as a computer device, a computer-readable storage medium, or a computer program product. Among them, the computer device includes a processor and a memory. The memory stores a computer program, and the processor executes the program to implement each step of the above-mentioned projection control method. The computer-readable storage medium stores computer instructions, and these instructions are used to make a computer execute the above-mentioned projection control method. The computer instructions included in the computer program product implement the above method steps when executed by the processor.

[0228] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment for details.

[0229] The above is only a preferred embodiment of the present application and is not used to limit the protection scope of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application are included in the protection scope of the present application.

[0230] The system of the embodiment of the present disclosure can execute the method provided by the embodiment of the present disclosure, and the implementation principle is similar. The actions performed by each module in the system of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed function description of each module of the system, reference can be specifically made to the description in the corresponding method shown above, and details will not be repeated here.

[0231] The above are only alternative implementation manners of some implementation scenarios of the present disclosure. It should be noted that for those of ordinary skill in the art, without departing from the technical concept of the solution of the present disclosure, other similar implementation means based on the technical idea of the present disclosure also fall within the protection scope of the embodiments of the present disclosure.

Claims

1. An intelligent adaptive projection control method, characterized in that: include: Collecting multi-source environmental data, wherein the multi-source environmental data includes at least one of environmental lighting data, projection surface characteristic data, and user operation data; Preprocessing, feature extraction and feature fusion of the multi-source environmental data to generate an environmental feature vector; Projecting the environment feature vector using a parameter prediction model based on rate-distortion optimization to generate projection parameters; Performing image deblurring processing and adaptive k-space sampling reconstruction on the projection parameters to generate an optimized projection image; Collecting projection effect feedback data for the optimized projection image; Performing online updating and optimization on the parameter prediction model according to the projection effect feedback data; The performing image deblurring processing and adaptive k-space sampling reconstruction on the projection parameters to generate an optimized projection image comprises: A neural network encoder based on prior knowledge encodes the projection image corresponding to the projection parameters to obtain an implicit representation of the feature encoding of the projection image; A decoder network based on a multi-scale structure restores the implicit representation of the feature encoding by layer-by-layer feature reconstruction to obtain a reconstructed image corresponding to the projected image; The reconstructed image is fused with the projection image through a residual learning strategy to generate a clear projection image; Based on a learnable sampling pattern generator, an adaptive sampling operation is performed according to the content characteristics of the clear projection image to determine the optimal sampling trajectory; A reconstruction model based on a deep convolutional network of the U-Net architecture is used to preserve image details of different scales through skip connections. Based on Fourier consistency constraints, the optimal sampling trajectory and image details of different scales are operated consistently to generate the initial optimized projection image; An image registration network is used to perform precise alignment, and geometric correction is performed on the initial optimized projection image to generate the optimized projection image.

2. The projection control method according to claim 1, characterized in that: The collecting of multi-source environmental data includes: Collecting the ambient light data by an ambient light sensor, wherein the ambient light data includes ambient light intensity data; Acquiring the projection surface characteristic data through a projection surface detector, wherein the projection surface characteristic data includes reflectivity and texture feature data of the projection surface; Through the user interaction interface, the user's adjustment operation data on the projection parameters is collected to obtain the user operation data.

3. The projection control method according to claim 1, characterized in that: The preprocessing, feature extraction and feature fusion of the multi-source environmental data to generate an environmental feature vector includes: Performing noise filtering and outlier detection processing on the ambient light data to obtain cleaned ambient light data; Performing geometric correction and texture feature extraction on the projection surface characteristic data to obtain standardized surface characteristic data; Performing time sequence alignment and behavior pattern analysis on the user operation data to obtain standardized user operation data; The convolutional neural network is used to extract the standardized surface feature data to obtain image features; Use recurrent neural networks to analyze normalized user operation data and obtain time series features; The image features, the time series features and the cleaned ambient light data are fused through an attention mechanism, and the fused data is the ambient feature vector.

4. The projection control method according to claim 1, characterized in that: The parameter prediction model is a multi-layer deep neural network, and the multi-layer deep neural network includes a feature receiving layer, a multi-layer fully connected layer based on a rectified linear unit activation function, and a parameter prediction layer; The step of performing projection mapping on the environmental feature vector to generate projection parameters by using a parameter prediction model based on rate distortion optimization includes: Receiving the environment feature vector through the feature receiving layer; In the fully connected layer, a deep representation corresponding to the environment feature vector is generated through a nonlinear transformation based on a ReLU activation function and a residual connection; In the parameter prediction layer, projection mapping is performed on the deep representation corresponding to the environmental feature vector based on multiple independent prediction units to generate the projection parameters, wherein one prediction unit corresponds to one projection parameter.

5. The projection control method according to claim 1 or 4, characterized in that: The parameter prediction model based on rate-distortion optimization is obtained in the following manner; Construct training data sets and validation sets based on pre-collected historical environmental data and corresponding optimal projection parameters; Constructing a loss function based on rate-distortion theory, wherein the loss function includes a parameter prediction error term and a rate constraint term; Using an improved stochastic gradient descent algorithm, based on the loss function and the training data set, to train the preset neural network model, and in the training process, using an adaptive moment estimation optimizer and a dynamic learning rate adjustment strategy; The model effect is evaluated based on the validation set, and hyperparameters including batch size and learning rate are dynamically adjusted to generate the parameter prediction model.

6. The projection control method according to claim 5, characterized in that: In the step of obtaining a reconstructed image corresponding to the projected image, the projection control method further includes: adopting a spatial attention mechanism to enhance the detail performance of important areas, and adding an adaptive weight adjustment strategy to dynamically adjust the deblurring strength according to the blur degree of the image area.

7. The projection control method according to claim 1, characterized in that: The collecting of projection effect feedback data for the optimized projection image includes: The brightness uniformity, color reproduction and contrast of the projected image are collected through optical sensors to form a time-series data stream; Determining quantitative indicators of the projection image by a preset evaluation algorithm, wherein the quantitative indicators include peak signal-to-noise ratio and structural similarity; Collecting user evaluation data through a user interaction interface, wherein the user evaluation data includes satisfaction ratings and feedback information; Record the user's parameter adjustment operation behavior data; The time series data stream, the quantitative index, the user evaluation data and the operation behavior data are cleaned and standardized to obtain standardized projection effect feedback data.

8. The projection control method according to claim 7, characterized in that: The online updating and optimization of the parameter prediction model according to the projection effect feedback data includes: Based on the normalized projection effect feedback data, an incremental learning strategy is adopted to gradually update the parameters of the parameter prediction model through a small batch gradient descent method, and in the parameter updating process, based on a dynamic learning rate adjustment mechanism, the parameter prediction model is adjusted according to the consistency of the normalized projection effect feedback data, and based on the knowledge distillation technology, the empirical knowledge of the historical model is transferred to the updated parameter prediction model; By comparing the projection qualities of the projection images before and after the parameter prediction model is updated, and based on the comparison result, verifying the performance improvement of the parameter prediction model, and deciding whether to accept the update based on the comparison result.

9. An intelligent adaptive projection lamp, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program, and when the processor executes the computer program, the projection control method according to any one of claims 1 to 8 is implemented.

Images