Backlight brightness control method and system based on reinforcement learning

Abstract

The invention discloses a backlight brightness control method and system based on reinforcement learning, and relates to the technical field of intelligent control. The method comprises the steps of performing state sensing on a target display area to obtain multiple pieces of state sensing data; extracting a plurality of display content images, and performing image semantic conversion to obtain a plurality of display content image semantic description texts; obtaining a plurality of fused semantic description features; updating the plurality of state sensing data, and analyzing backlight brightness control parameters to obtain a plurality of backlight brightness control parameters; and carrying out centralized screening analysis on the plurality of backlight source brightness control parameters to obtain a target backlight source brightness control parameter, and carrying out backlight source brightness control. The technical problem that in the prior art, backlight brightness adjustment depends on a fixed rule, semantic feature perception and self-adaptive control over the display content are lacked, and consequently the display quality and the energy-saving effect are difficult to consider at the same time is solved, and the technical effect of achieving intelligent brightness adjustment according to the display content through reinforcement learning is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control, in particular to a backlight brightness control method and system based on reinforcement learning. BACKGROUND

[0002] With the wide application of liquid crystal displays, vehicle-mounted display screens and mobile terminals, backlight brightness control has become an important link to improve display quality and reduce energy consumption. However, most of the existing backlight brightness adjustment methods still rely on fixed rules or single sensor feedback, and the adjustment results are usually only corrected for overall environmental brightness, lacking in-depth understanding of the differences in display content. In practical applications, display pictures often contain complex situations such as high and low brightness areas, static and dynamic scenes, etc., and it is difficult to balance the comprehensive needs of picture clarity, visual comfort and energy saving by relying on fixed rules for brightness adjustment. For example, when there are dark details in the picture, the traditional method is easy to cause detail loss, and in high-brightness scenes, the problem of high brightness and increased energy consumption may occur. In addition, the existing methods generally lack deep perception of semantic features and state data of display content, making the brightness control lack of intelligence and self-adaptive ability, resulting in insufficient control precision, which is difficult to meet the dual requirements of high-quality display and energy saving optimization. SUMMARY

[0003] The present application provides a backlight brightness control method and system based on reinforcement learning, which solves the technical problems that the backlight brightness adjustment in the prior art relies on fixed rules, lacks deep perception of semantic features and state data of display content, and results in insufficient brightness control precision, difficulty in balancing display quality and energy saving effect.

[0004] In a first aspect, the present application provides a backlight brightness control method based on reinforcement learning, which comprises: Using multi-agent reinforcement learning technology, calling multiple agents to perform state perception on a target display area according to a preset state index, obtaining multiple state perception data; extracting multiple display content images from the multiple state perception data, respectively performing image semantic conversion on the multiple display content images, and obtaining multiple display content image semantic description texts; performing directional feature fusion based on the multiple display content image semantic description texts, obtaining multiple fused semantic description features; updating the multiple state perception data according to the multiple fused semantic description features, and calling multiple agents to perform backlight brightness control parameter analysis on the multiple updated state perception data, obtaining multiple backlight brightness control parameters; centrally screening and analyzing the multiple backlight brightness control parameters, obtaining a target backlight brightness control parameter, and inputting the target backlight brightness control parameter into a backlight controller of the target display area for backlight brightness control.

[0005] In a second aspect, the present application provides a backlight brightness control system based on reinforcement learning, comprising: The state perception module adopts multi-agent reinforcement learning technology, calls multiple agents to perform state perception on the target display area according to preset state indicators, and obtains multiple state perception data; the semantic conversion module extracts multiple display content images from the multiple state perception data, respectively performs image semantic conversion on the multiple display content images, and obtains multiple display content image semantic description texts; the feature fusion module performs directional feature fusion based on the multiple display content image semantic description texts, and obtains multiple fused semantic description features; the parameter analysis module updates the multiple state perception data according to the multiple fused semantic description features, calls multiple agents to perform backlight brightness control parameter analysis on the multiple updated state perception data, and obtains multiple backlight brightness control parameters; and the brightness control module performs centralized screening analysis on the multiple backlight brightness control parameters, obtains target backlight brightness control parameters, and inputs the target backlight brightness control parameters into a backlight controller of the target display area to perform backlight brightness control.

[0006] The one or more technical solutions provided in the present application have at least the following technical effects or advantages: Firstly, the multi-agent reinforcement learning technology is adopted to call multiple agents to perform state perception on the target display area according to preset state indicators, and obtain multiple state perception data. Then, multiple display content images are extracted from the multiple state perception data, and image semantic conversion is respectively performed on the multiple display content images to obtain multiple display content image semantic description texts. Further, directional feature fusion is performed based on the multiple display content image semantic description texts to obtain multiple fused semantic description features. Then, the multiple state perception data are updated according to the multiple fused semantic description features, and multiple agents are called to perform backlight brightness control parameter analysis on the multiple updated state perception data to obtain multiple backlight brightness control parameters. Finally, centralized screening analysis is performed on the multiple backlight brightness control parameters to obtain target backlight brightness control parameters, and the target backlight brightness control parameters are input into a backlight controller of the target display area to perform backlight brightness control. The technical problems that the backlight brightness adjustment in the prior art relies on fixed rules, lacks deep perception of semantic features and state data of display content, and the brightness control precision is insufficient, and it is difficult to balance display quality and energy saving effect are solved. The technical effect that adaptive brightness control is achieved through multi-agent reinforcement learning and semantic feature fusion is achieved, so that the display picture clarity and visual comfort are improved while the energy consumption is effectively reduced. BRIEF DESCRIPTION OF DRAWINGS

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic flowchart of a backlight brightness control method based on reinforcement learning provided in an embodiment of this application; Figure 2 This is a schematic diagram of the backlight brightness control system based on reinforcement learning provided in an embodiment of this application.

[0009] Figure labeling: 11 State perception module, 12 Semantic conversion module, 13 Feature fusion module, 14 Parameter parsing module, 15 Brightness control module. Detailed Implementation

[0010] This application provides a backlight brightness control method and system based on reinforcement learning, which solves the technical problems in the prior art where backlight brightness adjustment relies on fixed rules and lacks deep perception of the semantic features and state data of the displayed content, resulting in insufficient brightness control accuracy and difficulty in balancing display quality and energy saving.

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0012] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0013] Example 1, as Figure 1 As shown, this application provides a backlight brightness control method based on reinforcement learning, wherein the method includes: By employing multi-agent reinforcement learning technology, multiple agents are invoked to perceive the state of the target display area according to preset state indicators, thereby obtaining multiple state perception data.

[0014] Furthermore, the preset status indicators include the brightness of the target display area, the ambient light intensity, the displayed content image, and the backlight energy efficiency.

[0015] Specifically, a plurality of state acquisition channels are arranged or virtually modeled in the target display area, each agent corresponding to an independent state acquisition channel, which can acquire the brightness distribution, ambient light intensity, display content image frame and backlight energy efficiency parameter of the target display area in real time, and perform feature quantization and normalization processing on the acquisition results to form a standardized state vector. Subsequently, a plurality of agents in the multi-agent reinforcement learning framework are called to perceive and feedback the above state vector, different agents performing policy exploration and policy utilization tasks, respectively, to correspondingly generate a plurality of state perception results. Each state perception result includes perception values of the real-time brightness level, local area light difference, display content feature strength and backlight energy consumption of the target display area. By aggregating the plurality of agent perception results, a plurality of state perception data are obtained.

[0016] Extracting a plurality of display content images from the plurality of state perception data, and performing image semantic conversion on the plurality of display content images respectively to obtain a plurality of display content image semantic description texts.

[0017] Further, extracting a plurality of display content images from the plurality of state perception data, and performing image semantic conversion on the plurality of display content images respectively to obtain a plurality of display content image semantic description texts, includes: Preconstructing an image semantic converter based on a convolutional neural network; and using the image semantic converter to perform image semantic conversion on the plurality of display content images to obtain the plurality of display content image semantic description texts.

[0018] The image semantic converter is preconstructed based on a convolutional neural network, and through step-by-step feature extraction of convolutional layers, pooling layers and fully connected layers, the image semantic converter can perform multi-level analysis of local features and global features of an input image, and in the training stage, supervised learning is performed using a labeled display content image sample set, so as to have the ability to identify text, icons, image subjects, background light characteristics and motion edges. A plurality of display content images are sequentially input into the image semantic converter to perform feature extraction and semantic mapping operations, and low-level pixel point information is gradually converted into high-level semantic feature vectors, and the semantic feature vectors are converted into corresponding semantic description texts in combination with a semantic dictionary generated in the training process. The semantic description text can represent the core semantic information of the display content in the form of natural language symbolization, such as “the image subject is text content, the background brightness is dark”, “the image subject is a high dynamic range video, containing a high-speed moving object”, “the image subject is a static icon, the edge region color saturation is high”, and the like.

[0019] Based on the plurality of display content image semantic description texts, directional feature fusion is performed to obtain a plurality of fused semantic description features.

[0020] Further, based on the plurality of display content image semantic description texts, directional feature fusion is performed to obtain a plurality of fused semantic description features, including: The semantic description features are extracted by traversing the plurality of display content image semantic description texts to obtain a plurality of semantic description features; the similarity of each pair of semantic description features is calculated to obtain an enumeration similarity set; the plurality of semantic description features are screened based on the enumeration similarity set to determine a leading semantic description feature and a plurality of following semantic description features; and the plurality of following semantic description features are directionally fused based on the leading semantic description feature to obtain the plurality of fused semantic description features.

[0021] Specifically, the feature of each semantic description text is decomposed by traversing the plurality of display content image semantic description texts to extract a corresponding semantic description feature vector to form a plurality of semantic description features; each pair of semantic description features is enumerated in a two-by-two combination manner, and the similarity of each pair of semantic description features is calculated. The similarity can be calculated by using cosine similarity, Euclidean distance, etc. Finally, an enumeration similarity set covering all feature pairs is obtained. The plurality of semantic description features are screened based on the enumeration similarity set, the average similarity of each feature with other features is calculated, the semantic description feature with the highest average similarity is determined as the leading semantic description feature, and the remaining semantic description features are marked as following semantic description features. The leading semantic description feature is taken as the center to construct a directional fusion matrix, and the plurality of following semantic description features are subjected to weighted directional fusion processing. The semantic leading direction of the leading feature is maintained during the fusion process, and a multi-dimensional semantic vector representation after fusion is output by iteration and superposition. Finally, a plurality of fused semantic description features are obtained.

[0022] Further, based on the enumeration similarity set, the plurality of semantic description features are screened to determine a leading semantic description feature and a plurality of following semantic description features, including: The enumeration similarity set is searched with the plurality of semantic description features as indexes to obtain a plurality of semantic description feature associated enumeration similarity sets; the mean of the plurality of semantic description feature associated enumeration similarity sets is calculated to obtain a plurality of semantic description feature associated enumeration similarity means; the maximum value in the plurality of semantic description feature associated enumeration similarity means is taken as the leading semantic description feature, and the remaining semantic description features are taken as the plurality of following semantic description features.

[0023] Firstly, the enumeration similarity set is indexed according to the semantic description features, and each semantic description feature is taken as a retrieval key to extract the similarity values between the semantic description feature and all other semantic description features, thereby forming a plurality of enumeration similarity sets associated with the semantic description features. Subsequently, the enumeration similarity sets associated with each semantic description feature are subjected to mean value calculation to obtain the enumeration similarity mean values associated with the semantic description features, thereby reflecting the centrality and representativeness of the semantic description features in the overall set. Next, the enumeration similarity mean values associated with the plurality of semantic description features are compared, and the one with the largest value is selected as the leading semantic description feature to represent the most representative and directional feature in the global semantic space. Finally, the remaining semantic description features are sequentially marked as following semantic description features in the order of mean value.

[0024] Further, the plurality of following semantic description features are directionally fused based on the leading semantic description feature to obtain the plurality of fused semantic description features, including: The plurality of following semantic description features are subjected to fine-grained sub-feature similarity identification based on the leading semantic description feature to obtain a plurality of following fine-grained sub-feature similarity sets; the plurality of directionality leading matrices are constructed by traversing and preprocessing the plurality of following fine-grained sub-feature similarity sets; and the plurality of following semantic description features are directionally fused based on the plurality of directionality leading matrices to obtain the plurality of fused semantic description features in combination with the leading semantic description feature.

[0025] Further, the plurality of following fine-grained sub-feature similarity sets are subjected to normalization processing, and the plurality of normalized processing results are filled into the initially empty matrices to obtain the plurality of directionality leading matrices.

[0026] Specifically, each following semantic description feature is divided into fine-grained sub-features with reference to the leading semantic description feature, and the semantic vector is decomposed into a plurality of sub-feature units according to the dimensions; subsequently, the similarity between each sub-feature unit and the corresponding sub-feature of the leading semantic description feature is calculated one by one to obtain a plurality of following fine-grained sub-feature similarity sets, and the similarity can be measured by methods such as cosine similarity or Euclidean distance. Then, the plurality of following fine-grained sub-feature similarity sets are traversed and preprocessed, the similarity values in each set are normalized, and the normalized results are filled into the initially empty matrices to construct a plurality of directionality leading matrices. Finally, the plurality of following semantic description features are directionally fused with weighting under the constraint of the plurality of directionality leading matrices, the leading semantic description feature maintains the dominant direction in the fusion process, and the following features are subjected to similarity weighting and superposition, thereby forming the high-level semantic feature representation after fusion, which is output in combination with the leading semantic description feature to finally obtain the plurality of fused semantic description features.

[0027] The plurality of state perception data is updated according to the plurality of fusion semantic description features, and a plurality of agents are called to parse the plurality of updated state perception data to obtain a plurality of backlight brightness control parameters.

[0028] Specifically, the plurality of fusion semantic description features obtained by fusing the directional features are correspondingly mapped with the plurality of state perception data, and the brightness parameter, the ambient light parameter, the display content feature parameter and the backlight energy efficiency parameter in each state perception data are weighted and updated. The weighting factor is adaptively adjusted according to the similarity between the fusion semantic description feature and the original feature, so as to generate a plurality of updated state perception data. Subsequently, the plurality of updated state perception data is input into a multi-agent reinforcement learning framework, and each agent parses the input data based on its own policy network to output a corresponding backlight brightness control parameter candidate value. Specifically, the agent generates control parameters including brightness adjustment amplitude, backlight power distribution ratio and energy efficiency constraint factor through a state-action-reward cycle mechanism combined with display content type and ambient light intensity. Finally, the parsing results of all agents are collected to obtain a plurality of backlight brightness control parameters.

[0029] The plurality of backlight brightness control parameters are centrally screened and analyzed to obtain a target backlight brightness control parameter, and the target backlight brightness control parameter is input into a backlight controller of the target display area for backlight brightness control.

[0030] The plurality of backlight brightness control parameters are centrally screened and analyzed to obtain a target backlight brightness control parameter; the target backlight brightness control parameter is input into a backlight controller of the target display area, and the backlight controller performs adjustment operation of driving current and voltage according to the target backlight brightness control parameter to control backlight brightness output in real time, so as to realize dynamic adaptive adjustment of backlight brightness under different display content and ambient light conditions.

[0031] Further, the plurality of backlight brightness control parameters are centrally screened and analyzed to obtain a target backlight brightness control parameter, including: The plurality of backlight brightness control parameters are centrally screened and analyzed to obtain a target backlight brightness control parameter; the target backlight brightness control parameter is input into a backlight controller of the target display area, and the backlight controller performs adjustment operation of driving current and voltage according to the target backlight brightness control parameter to control backlight brightness output in real time, so as to realize dynamic adaptive adjustment of backlight brightness under different display content and ambient light conditions.

[0032] Further, the preset screening condition is that the number of central screening is less than or equal to a preset number threshold.

[0033] Specifically, the backlight brightness control parameters output by all the multiple agents are statistically processed to calculate the arithmetic mean value of the backlight brightness control parameters, and the mean value is taken as the initial reference point for centralized screening. Taking the mean value of the backlight brightness control parameters as the starting point, the screening range is updated iteratively in the parameter set according to the preset centralized screening step size. At each iteration, the parameters deviating from the current reference value by more than the preset tolerance threshold are removed, and the mean value of the remaining parameters is recalculated as a new reference point. In this way, the parameter set is gradually shrunk until the screening process meets the preset screening condition, that is, when the number of centralized screening reaches the preset number threshold, the further screening operation is immediately terminated, and the parameter set obtained by the last iteration screening is taken as the final result, and the target backlight brightness control parameter is output.

[0034] To sum up, the embodiments of the present application have at least the following technical effects: First, the multi-agent reinforcement learning technology is used to call multiple agents to perform state sensing on the target display area according to the preset state indicators, and obtain multiple state sensing data. Then, multiple display content images in the multiple state sensing data are extracted, and image semantic conversion is performed on the multiple display content images respectively to obtain multiple display content image semantic description texts. Further, directional feature fusion is performed based on the multiple display content image semantic description texts to obtain multiple fused semantic description features. Then, the multiple state sensing data are updated according to the multiple fused semantic description features, and the multiple agents are called to analyze the multiple updated state sensing data to obtain multiple backlight brightness control parameters. Finally, the target backlight brightness control parameter is obtained by centralized screening analysis of the multiple backlight brightness control parameters, and the target backlight brightness control parameter is input into the backlight controller of the target display area for backlight brightness control. The technical problem that the backlight brightness adjustment in the prior art relies on fixed rules and lacks deep sensing of the semantic features and state data of the display content, resulting in insufficient brightness control precision and difficulty in balancing display quality and energy saving effect is solved. The technical effect of achieving adaptive brightness control through multi-agent reinforcement learning and semantic feature fusion is achieved, so that the display picture clarity and visual comfort are improved while the energy consumption is effectively reduced.

[0035] Embodiment two, based on the same inventive concept as the backlight brightness control method based on reinforcement learning in the foregoing embodiments, as shown in Figure 2 The present application provides a backlight brightness control system based on reinforcement learning, wherein the system comprises: The state perception module 11: adopts multi-agent reinforcement learning technology, calls multiple agents to perform state perception on the target display area according to a preset state index, and obtains multiple state perception data; the semantic conversion module 12: extracts multiple display content images in the multiple state perception data, respectively performs image semantic conversion on the multiple display content images, and obtains multiple display content image semantic description texts; the feature fusion module 13: performs directional feature fusion based on the multiple display content image semantic description texts, and obtains multiple fusion semantic description features; the parameter analysis module 14: updates the multiple state perception data according to the multiple fusion semantic description features, and calls multiple agents to perform backlight brightness control parameter analysis on the multiple updated state perception data, and obtains multiple backlight brightness control parameters; the brightness control module 15: centrally screens and analyzes the multiple backlight brightness control parameters, obtains a target backlight brightness control parameter, and inputs the target backlight brightness control parameter into a backlight controller of the target display area to perform backlight brightness control.

[0036] Further, the state perception module 11 is configured to perform the following method: The preset state index includes target display area brightness, ambient light intensity, display content image, and backlight energy efficiency.

[0037] Further, the semantic conversion module 12 is configured to perform the following method: Pre-construct an image semantic converter based on a convolutional neural network; use the image semantic converter to perform image semantic conversion on the multiple display content images, and obtain the multiple display content image semantic description texts.

[0038] Further, the feature fusion module 13 is configured to perform the following method: Iterate the multiple display content image semantic description texts to extract semantic description features, obtain multiple semantic description features; calculate the similarity of each two of the multiple semantic description features, obtain an enumeration similarity set; filter the multiple semantic description features based on the enumeration similarity set, determine a leading semantic description feature and multiple following semantic description features; perform directional fusion on the multiple following semantic description features based on the leading semantic description feature, and obtain the multiple fusion semantic description features.

[0039] Further, the feature fusion module 13 is configured to perform the following method: The enumeration similarity set is searched respectively with the plurality of semantic description features as indexes, a plurality of semantic description feature associated enumeration similarity sets are obtained, a mean value of the plurality of semantic description feature associated enumeration similarity sets is calculated, and a plurality of semantic description feature associated enumeration similarity mean values are obtained; a maximum value in the plurality of semantic description feature associated enumeration similarity mean values is taken, and a corresponding semantic description feature is taken as the leading semantic description feature, and the remaining semantic description features are taken as a plurality of following semantic description features.

[0040] Further, the feature fusion module 13 is configured to perform the following method: Based on the leading semantic description feature, fine-grained sub-feature similarity recognition is performed on the plurality of following semantic description features, a plurality of following fine-grained sub-feature similarity sets are obtained, the plurality of following fine-grained sub-feature similarity sets are preprocessed, a plurality of directional leading matrices are constructed, and based on the plurality of directional leading matrices, directional fusion is performed on the plurality of following semantic description features, and in combination with the leading semantic description feature, the plurality of fused semantic description features are obtained.

[0041] Further, the feature fusion module 13 is configured to perform the following method: The plurality of following fine-grained sub-feature similarity sets are normalized respectively, and a plurality of normalized processing results are filled into an initially empty matrix respectively, and the plurality of directional leading matrices are obtained.

[0042] Further, the brightness control module 15 is configured to perform the following method: The plurality of backlight brightness control parameters are mean calculated, and a backlight brightness control parameter mean value is obtained; the backlight brightness control parameter mean value is taken as a starting point, a plurality of backlight brightness control parameters are centrally screened according to a preset central screening step, until a preset screening condition is met, and a result obtained by the last screening is taken as the target backlight brightness control parameter.

[0043] Further, the brightness control module 15 is configured to perform the following method: The preset screening condition is that the number of central screening is less than or equal to a preset number threshold.

[0044] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above-mentioned specific embodiments of the present application are described. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0045] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0046] The specification and drawings are only exemplary and illustrative of the present application and are considered to cover any and all modifications, variations, combinations or equivalents that are within the scope of the present application. Obviously, various modifications and changes can be made to the present application by those skilled in the art without departing from the scope of the present application. Thus, it is intended that the present application cover the modifications and changes as they come within the scope of the application, and that the scope of the application be defined not by the description of the application, but by the claims.

Claims

1. A backlight brightness control method based on reinforcement learning, characterized in that, The method includes: By employing multi-agent reinforcement learning technology, multiple agents are invoked to perceive the state of the target display area according to preset state indicators, thereby obtaining multiple state perception data. Extract multiple display content images from the multiple state-aware data, and perform image semantic transformation on the multiple display content images respectively to obtain semantic description text of multiple display content images; Based on the semantic description text of the multiple displayed content images, directional features are fused to obtain multiple fused semantic description features; The multiple state-aware data are updated based on the multiple fused semantic description features, and multiple intelligent agents are invoked to parse the backlight brightness control parameters of the multiple updated state-aware data to obtain multiple backlight brightness control parameters. The multiple backlight brightness control parameters are centrally screened and analyzed to obtain the target backlight brightness control parameters, which are then input into the backlight controller of the target display area for backlight brightness control.

2. The backlight brightness control method based on reinforcement learning as described in claim 1, characterized in that, The preset status indicators include the brightness of the target display area, the ambient light intensity, the displayed content image, and the backlight energy efficiency.

3. The backlight brightness control method based on reinforcement learning as described in claim 1, characterized in that, Extract multiple display content images from the multiple state-aware data, and perform image semantic transformation on each of the multiple display content images to obtain semantic description text for the multiple display content images, including: Pre-built image semantic converter based on convolutional neural network; The image semantic converter is used to perform image semantic conversion on the plurality of display content images to obtain semantic description text for the plurality of display content images.

4. The backlight brightness control method based on reinforcement learning as described in claim 1, characterized in that, Based on the semantic description text of the multiple displayed content images, directional features are fused to obtain multiple fused semantic description features, including: The semantic description features are extracted by traversing the semantic description text of the multiple displayed content images to obtain multiple semantic description features; The pairwise similarity calculation is performed on the multiple semantic description features to obtain an enumerated similarity set; Based on the enumerated similarity set, the multiple semantic description features are filtered to determine the leading semantic description feature and multiple following semantic description features; Based on the leading semantic description feature, the multiple following semantic description features are directionally fused to obtain the multiple fused semantic description features.

5. The backlight brightness control method based on reinforcement learning as described in claim 4, characterized in that, Based on the enumerated similarity set, the multiple semantic description features are filtered to determine the leading semantic description feature and multiple following semantic description features, including: Using the multiple semantic description features as indexes, the enumerated similarity set is retrieved to obtain multiple semantic description feature associated enumerated similarity sets; Calculate the mean of the multiple semantic description feature association enumeration similarity sets to obtain the mean of the multiple semantic description feature association enumeration similarity; The maximum value among the average similarity values ​​of the multiple semantic description features is taken as the leading semantic description feature, and the remaining semantic description features are taken as multiple following semantic description features.

6. The backlight brightness control method based on reinforcement learning as described in claim 5, characterized in that, Based on the leading semantic description feature, the multiple following semantic description features are directionally fused to obtain the multiple fused semantic description features, including: Based on the leading semantic description features, fine-grained sub-feature similarity recognition is performed on the multiple following semantic description features to obtain multiple sets of following fine-grained sub-feature similarities; The multiple sets of similarity to follow-fine-grained sub-features are traversed for preprocessing to construct multiple directional guiding matrices; Based on the multiple directional guiding matrices, the multiple following semantic description features are directionally fused, and combined with the guiding semantic description features, the multiple fused semantic description features are obtained.

7. The backlight brightness control method based on reinforcement learning as described in claim 6, characterized in that, The multiple sets of similarity of the following fine-grained sub-features are normalized respectively, and the results of the multiple normalization processes are filled into the initially empty matrix to obtain the multiple directional guidance matrices.

8. The backlight brightness control method based on reinforcement learning as described in claim 1, characterized in that, The target backlight brightness control parameters are obtained by performing centralized screening and analysis on the multiple backlight brightness control parameters, including: The average value of the backlight brightness control parameters is calculated by averaging the multiple backlight brightness control parameters. Starting from the average value of the backlight brightness control parameters, multiple backlight brightness control parameters are centrally screened according to a preset central screening step size until the preset screening conditions are met. The result obtained from the last screening is then used as the target backlight brightness control parameter.

9. The backlight brightness control method based on reinforcement learning as described in claim 8, characterized in that, The preset filtering condition is that the number of times the data is filtered is less than or equal to the preset number of times threshold.

10. A backlight brightness control system based on reinforcement learning, characterized in that, The system is used to implement the backlight brightness control method based on reinforcement learning as described in any one of claims 1-9, the system comprising: State perception module: Employs multi-agent reinforcement learning technology, calling multiple agents to perceive the state of the target display area according to preset state indicators, and obtaining multiple state perception data; Semantic conversion module: Extracts multiple display content images from the multiple state-aware data, performs image semantic conversion on the multiple display content images respectively, and obtains semantic description text of multiple display content images; Feature fusion module: Performs directional feature fusion based on the semantic description text of the multiple displayed content images to obtain multiple fused semantic description features; Parameter parsing module: Updates the multiple state perception data according to the multiple fused semantic description features, and calls multiple intelligent agents to parse the backlight brightness control parameters of the multiple updated state perception data to obtain multiple backlight brightness control parameters; Brightness control module: performs centralized screening and analysis on the multiple backlight brightness control parameters to obtain target backlight brightness control parameters, and inputs the target backlight brightness control parameters into the backlight controller of the target display area for backlight brightness control.

Images