An artificial intelligence image recognition system based on reinforcement learning

Abstract

The application discloses an image recognition system based on reinforcement learning for artificial intelligence, and relates to the technical field of artificial intelligence image recognition, comprising a multi-device image input module, an optimal transmission color alignment preprocessing module, a Q-learning convolution kernel dynamic configuration feature extraction module, an enhanced learning recognition decision module, a performance monitoring module, an adaptive adjustment module and a self-optimization module, wherein the multi-device image input module is used for receiving image data from multi-source heterogeneous devices and detecting and identifying the device type and color space type corresponding to the image data; the application has the advantages that the optimal transmission color alignment preprocessing technology is adopted to fundamentally solve the cross-device color distribution mismatch problem, the image input of multi-source heterogeneous devices such as monitoring cameras, industrial detection cameras, consumer-grade mobile phones and remote sensing satellites can be compatible, the cross-device recognition accuracy is improved compared with the existing normalization scheme, and the system versatility is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence image recognition technology, specifically to an image recognition system for artificial intelligence based on reinforcement learning. Background Technology

[0002] Image recognition technology based on reinforcement learning is a current research hotspot in the field of artificial intelligence. It adopts the Actor-Critic framework to optimize the update strategy of the classifier for small sample incremental scenarios, which solves the catastrophic forgetting problem in model training. However, this type of technology is only suitable for standard natural images with fixed resolution. The convolution kernel parameters of the backbone network are completely fixed. The optimization scope of reinforcement learning is limited to the classifier module and cannot work effectively in image input scenarios of multi-source heterogeneous devices. The problem of color distribution mismatch across devices arises because the spectral response and color space of sensors on different devices are fundamentally different. This results in the same target appearing with completely different color characteristics on different devices. Conventional normalization preprocessing methods can only achieve resolution interpolation adjustment and cannot align the color distribution differences between different devices, leading to a significant decrease in cross-device recognition accuracy. Applying the optimal transmission algorithm to color alignment in the preprocessing stage is a rarely used technique in the industry, and none of the three existing representative technologies cover this solution. Therefore, we propose an image recognition system for artificial intelligence based on reinforcement learning. Summary of the Invention

[0003] The purpose of this invention is to provide an image recognition system for artificial intelligence based on reinforcement learning.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an image recognition system for artificial intelligence based on reinforcement learning, the system comprising a multi-device image input module, an optimal transmission color alignment preprocessing module, a Q-learning convolution kernel dynamic configuration feature extraction module, a reinforcement learning recognition decision module, a performance monitoring module, an adaptive adjustment module, and a self-optimization module; The multi-device image input module is used to receive image data from multiple heterogeneous devices and to detect and identify the device type and color space type corresponding to the image data. The optimal transmission color alignment preprocessing module is connected to the multi-device image input module. It is used to extract the RGB three-channel color distribution histogram of the input image, construct the optimal transmission cost matrix between the source color distribution and the standard color distribution based on the color distribution histogram, solve the optimal transmission cost matrix using the Sinkhorn-Knopp algorithm to obtain the optimal transmission matrix, and map and align the color distribution of the input image to the standard device color distribution based on the optimal transmission matrix, and output the color normalized preprocessed image. The Q-learning convolutional kernel dynamic configuration feature extraction module, connected to the optimal transmission color alignment preprocessing module, includes a first Q-learning agent. The first Q-learning agent uses the entropy value of the current input image and the target density to construct state features, and uses discrete combinations of the number and size of convolutional kernels to construct an action space. The action space includes three sets of discrete convolutional kernel configurations: the first set has 32 convolutional kernels and a kernel size of 7×7, corresponding to a coarse-grained feature extraction mode; the second set has 64 convolutional kernels and a kernel size of 3×3, corresponding to a medium-grained feature extraction mode; and the third set has 128 convolutional kernels and a kernel size of 1×1, corresponding to a fine-grained feature extraction mode. The first Q-learning agent adaptively selects a set of convolutional kernel configurations based on the state features of the current image to extract features from the color-normalized preprocessed image, outputting an image feature vector. The reinforcement learning recognition decision module is connected to the Q-learning convolution kernel dynamic configuration feature extraction module, and is used to make recognition decisions based on the image feature vector and output the image recognition result; The performance monitoring module is connected to the reinforcement learning recognition decision module. It is used to collect the current recognition accuracy index, computation latency index and memory usage index of the system, and to combine the three indexes with the current input resolution, image entropy value and feature dimension to form a five-dimensional state vector, which is used as the state input of the adaptive adjustment module. The adaptive adjustment module, connected to the performance monitoring module, includes a second Q-learning agent. The second Q-learning agent takes the five-dimensional state vector as input and a three-dimensional action space composed of preprocessing parameter adjustments, convolution kernel parameter adjustments, and learning rate adjustments as output. It calculates the reward value of the current action and updates the action policy based on a multi-objective reward function. The multi-objective reward function is the sum of three terms: a positive weighted value for recognition accuracy, a negative weighted value for computation latency, and a negative weighted value for memory usage. The weight coefficients corresponding to these three terms are adaptively updated by the second Q-learning agent during iteration according to the Pareto optimality principle. The self-optimization module is connected to the adaptive adjustment module, the optimal transmission color alignment preprocessing module, the Q-learning convolutional kernel dynamic configuration feature extraction module, and the reinforcement learning recognition decision module, respectively. It is used to distribute the three-dimensional action parameters output by the adaptive adjustment module to the optimal transmission color alignment preprocessing module, the Q-learning convolutional kernel dynamic configuration feature extraction module, and the reinforcement learning recognition decision module to complete the multi-objective closed-loop iterative update of the parameters of the entire system. The performance monitoring module, the adaptive adjustment module, and the self-optimization module together constitute the multi-objective Q-learning closed-loop framework of the whole system, realizing continuous self-optimization of the optimal transmission color alignment preprocessing module, the Q-learning convolution kernel dynamic configuration feature extraction module, and the reinforcement learning recognition decision module.

[0005] As a further aspect of the present invention: the elements of the optimal transmission cost matrix are determined according to the following formula: ; in, The first in the source color distribution The color value and the target color distribution in the _ ... The cost of transmitting individual color values The first in the source color distribution These are color values, specifically the color values ​​extracted from the image input to the device. For the target color distribution, the first Each color value corresponds to a color value in a standard device image. for and The cosine similarity between two color values ​​indicates that their hues are more similar, resulting in lower transmission costs. The first in the source color distribution The positional weight corresponding to each color value For the target color distribution, the first The positional weight corresponding to each color value This is a spatial regularization coefficient used to adjust the relative weight of color similarity and spatial distance in transmission cost calculations. The system pre-calibrates based on the device type.

[0006] As a further aspect of the present invention: the multi-objective reward function is determined according to the following formula: ; in, This represents the multi-objective reward value at the current moment. The normalized recognition accuracy of the current recognition result is set to a value of [value]. Real numbers within the interval; The computational delay for the current inference process, in milliseconds. The maximum acceptable inference latency threshold preset for the system, in milliseconds. This represents the memory usage of the current inference process, in megabytes. This represents the maximum available memory capacity of the system, in megabytes. and All The normalized dimensionless number within the interval, and Uniformity of dimensions ensures the mathematical validity of the weighted sum of the three terms. The Pareto weight coefficient corresponding to the recognition accuracy represents the proportion of positive contribution of recognition accuracy to the reward function; The Pareto weight coefficient corresponding to the delay is used to calculate the proportion of the negative penalty for the delay in the reward function. The Pareto weight coefficient corresponding to memory usage represents the proportion of negative penalty imposed on memory usage in the reward function; , , The second Q-learning agent adaptively adjusts itself in each closed-loop iteration based on the Pareto optimality principle.

[0007] As a further aspect of the present invention: the five dimensions of the five-dimensional state vector are, in order: the resolution of the current input image, the entropy value of the current input image, the number of feature dimensions of the current feature extraction layer, the recognition accuracy of the current system, and the computational delay of the current system.

[0008] As a further aspect of the present invention: the self-optimization module triggers a closed-loop iterative update when the system has processed a total of 1,000 images; after triggering, the performance monitoring module calculates the average recognition accuracy, average computation latency, and average memory usage of the 1,000 images in that batch; the adaptive adjustment module adjusts the weight coefficients according to the Pareto optimality principle based on the statistical results; the self-optimization module synchronously issues parameter update instructions based on the updated weight coefficients, respectively updating the preprocessing parameters of the optimal transmission color alignment preprocessing module, the convolution kernel parameters of the Q-learning convolution kernel dynamic configuration feature extraction module, and the learning rate parameters of the reinforcement learning recognition decision module.

[0009] As a further aspect of the present invention: the color space type identified by the multi-device image input module includes the sRGB color space of consumer-grade devices and the multispectral color space of remote sensing satellite devices; when the input image is detected to belong to the sRGB color space, the optimal transmission color alignment preprocessing module simultaneously activates downsampling preprocessing to reduce the amount of computation; when the input image is detected to belong to the multispectral color space, the optimal transmission color alignment preprocessing module simultaneously activates contrast-limited histogram equalization to improve image quality.

[0010] As a further aspect of the present invention: when the entropy value of the current image is lower than 0.5 and the target density is lower than 0.3, the first Q-learning agent selects the first set of convolutional kernel configurations to perform coarse-grained feature extraction; when the entropy value of the current image is between 0.5 and 0.7 and the target density is between 0.3 and 0.5, the first agent selects the second set of convolutional kernel configurations to perform medium-grained feature extraction; when the entropy value of the current image is higher than 0.7 and the target density is higher than 0.5, the first agent selects the third set of convolutional kernel configurations to perform fine-grained feature extraction.

[0011] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows: 1. This invention adopts optimal transmission color alignment preprocessing technology, which fundamentally solves the problem of color distribution mismatch across devices. It is compatible with image input from multiple heterogeneous devices such as surveillance cameras, industrial inspection cameras, consumer mobile phones, and remote sensing satellites. The cross-device recognition accuracy is improved compared with the existing normalization scheme, and the system versatility is greatly improved.

[0012] 2. This invention adopts Q-learning convolution kernel dynamic configuration technology to realize end-to-end intelligent decision-making of convolution kernel parameters, completely replacing manual preset rules, improving computational efficiency in simple scenarios and recognition accuracy in complex scenarios, and achieving a dynamic adaptive balance between accuracy and efficiency.

[0013] 3. This invention constructs a closed-loop framework for multi-objective Q-learning that integrates Pareto optimization, expanding the optimization scope of reinforcement learning from a single objective of a single module to multi-objective collaborative optimization of the entire system. This results in long-term iterative performance improvement of the system, enabling continuous performance self-upgrades without manual intervention and significantly reducing system maintenance costs. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the system flow in an embodiment of the present invention; Figure 2 This is a flowchart of the algorithm for the transmission color alignment preprocessing module in an embodiment of the present invention; Figure 3 This is an interactive flowchart of the multi-objective Q-learning closed-loop framework in an embodiment of the present invention. Detailed Implementation

[0015] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0016] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] Please see the appendix Figure 1 - Appendix Figure 3 The present invention discloses an image recognition system for artificial intelligence based on reinforcement learning, comprising a multi-device image input module, an optimal transmission color alignment preprocessing module, a Q-learning convolution kernel dynamic configuration feature extraction module, a reinforcement learning recognition decision module, a performance monitoring module, an adaptive adjustment module, and a self-optimization module.

[0018] The multi-device image input module is used to receive image data from multiple heterogeneous devices and to detect and identify the device type and color space type.

[0019] The optimal transmission color alignment preprocessing module is connected to the multi-device image input module. This module first extracts the RGB three-channel color distribution histogram of the input image, and constructs an optimal transmission cost matrix between the source color distribution and the standard color distribution based on the histogram data. The elements of the cost matrix... Calculate according to the following formula: ; In the formula, The first in the source color distribution The color value and the target color distribution in the _ ... The transmission cost between color values ​​is indicated by a smaller value, which represents the lower the cost of color mapping between the two values. The first in the source color distribution Each color value is a color sample point from the image input from the device. For the target color distribution, the first Each color value represents a color sampling point corresponding to a standard device image. for and The cosine similarity between two colors indicates that their hues are closer and the transmission cost is lower. The first in the source color distribution The positional weights corresponding to each color value reflect the spatial distribution contribution of that color value in the color histogram. For the target color distribution, the first The positional weight corresponding to each color value This is a spatial regularization coefficient used to adjust the relative weight of color cosine similarity and location distance in transmission cost calculation. The system is pre-calibrated according to the device type to adapt to the color distribution characteristics of different devices.

[0020] After constructing the cost matrix, the module uses the Sinkhorn-Knopp algorithm to iteratively solve the optimal transmission cost matrix to obtain the optimal transmission matrix. Based on the transmission matrix, the color distribution of the input image is mapped and aligned to the standard device color distribution, and the color-normalized preprocessed image is output.

[0021] The Q-learning convolutional kernel dynamic configuration feature extraction module is connected to the optimal transmission color alignment preprocessing module, and includes a first Q-learning agent. The first Q-learning agent uses the entropy value and target density of the current input image as two-dimensional state features to quantitatively evaluate the complexity of the image. The entropy value reflects the uniformity of the image pixel grayscale distribution; a higher entropy value indicates richer image information and a more complex scene. The target density reflects the density of target objects in the image; a higher target density indicates more recognition targets in the image and greater difficulty in feature extraction. The action space of the first Q-learning agent is composed of discrete combinations of the number and size of convolutional kernels, including three configurations: the first configuration has 32 convolutional kernels and a kernel size of 7×7, with a large receptive field and low computational cost, suitable for coarse-grained feature extraction; the second configuration has 64 convolutional kernels and a kernel size of 3×3, with a balanced receptive field and computational cost, suitable for medium-grained feature extraction; and the third configuration has 128 convolutional kernels. The convolutional kernel size is 1×1, which has strong channel feature extraction capability and high computational density, making it suitable for fine-grained feature extraction. The first Q-learning agent selects the first configuration when the current image entropy value is below 0.5 and the target density is below 0.3; selects the second configuration when the entropy value is between 0.5 and 0.7 and the target density is between 0.3 and 0.5; and selects the third configuration when the entropy value is above 0.7 and the target density is above 0.5. The agent continuously updates the Q-table through the Q-learning strategy to achieve self-optimization of the convolutional kernel configuration selection strategy. After feature extraction of the color-normalized preprocessed image with the selected convolutional kernel configuration, the module outputs the image feature vector. The target density is only used in the feature extraction configuration decision of a single image. Its calculation is completed by the optimal transmission color alignment preprocessing module in the preprocessing stage and passed to the first Q-learning agent. The second Q-learning agent performs cross-batch statistical optimization for the entire system. The target density is implicitly included in the entropy value index at the batch statistical level, so it is not listed separately in the state vector.

[0022] The reinforcement learning recognition decision module is connected to the Q-learning convolutional kernel dynamic configuration feature extraction module, and internally contains a recognition policy agent based on a deep Q-network (DQN): the state space of the recognition policy agent is a normalized representation of the image feature vector, the action space is a set of image category labels (K categories in total, K is determined by the specific task), and the reward function is defined as a reward of +1 when the recognition is correct and a reward of -1 when the recognition is incorrect; the agent iteratively updates the Q value. To achieve self-optimization of the recognition strategy, among which For the system's Q-learning agent in the current state Take action below The action value function is used to evaluate the expected cumulative reward of the current state action pair. Predicting the next state for a Q-learning agent and take action , The learning rate is dynamically adjusted by the self-optimization module. The discount factor is used; during the inference phase, the agent uses a greedy strategy to select the category label with the largest Q value as the recognition result output, makes recognition decisions based on the image feature vector, and outputs the image recognition result.

[0023] The performance monitoring module is connected to the reinforcement learning recognition decision module, and collects the system's current recognition accuracy, computation latency, and memory usage in real time. These three metrics, along with the current input resolution, image entropy, and feature dimension, together form a five-dimensional state vector, denoted as... ,in For the input image resolution, Image entropy value, The number of feature dimensions in the feature extraction layer. For the current recognition accuracy, The five-dimensional state vector serves as the state input for the adaptive adjustment module, representing the current computational delay.

[0024] The adaptive adjustment module is connected to the performance monitoring module and includes a second Q-learning agent. The second Q-learning agent takes the five-dimensional state vector as input and adjusts the preprocessing parameters accordingly. Kernel parameter adjustment and learning rate adjustment Constructing a three-dimensional action space For output, the second Q-learning agent calculates the reward value of the current action and updates the action policy based on a multi-objective reward function, the expression of which is: ; In the formula, This represents the multi-objective reward value at the current moment, used to measure the overall performance of the current system parameter configuration. The normalized recognition accuracy of the current recognition result is set to a value of [value]. For real numbers within the interval, the larger the value, the better the recognition effect. This represents the computational delay of the current inference process, measured in milliseconds. A smaller value indicates a faster inference speed. This represents the memory usage of the current inference process, in megabytes. A smaller value indicates higher memory utilization efficiency. The Pareto weight coefficient corresponding to the recognition accuracy represents the proportion of the positive contribution of recognition accuracy to the reward function. To calculate the Pareto weight coefficient corresponding to the delay, which represents the negative penalty of the delay in the reward function, This represents the Pareto weighting coefficient corresponding to memory usage, indicating the proportion of negative penalty imposed on memory usage in the reward function. , , The second Q-learning agent adaptively adjusts these three factors in each closed-loop iteration according to the Pareto optimality principle in order to find the Pareto optimal balance between recognition accuracy, computational latency, and memory usage.

[0025] The self-optimization module is connected to the adaptive adjustment module, the optimal transmission color alignment preprocessing module, the Q-learning convolutional kernel dynamic configuration feature extraction module, and the reinforcement learning recognition decision module, respectively, and transmits the three-dimensional action parameters output by the adaptive adjustment module. Distribute the parameters to each target module and simultaneously update the preprocessing parameters, convolution kernel parameters, and learning rate parameters to achieve multi-objective closed-loop iteration of the parameters of the entire system.

[0026] The performance monitoring module, the adaptive adjustment module, and the self-optimization module together constitute the multi-objective Q-learning closed-loop framework of the entire system. When the cumulative number of processed images reaches 1000, the system triggers a closed-loop iterative update: the performance monitoring module calculates the average recognition accuracy, average computation latency, and average memory usage of this batch; the adaptive adjustment module adjusts the Pareto weight coefficients based on the statistical results; and the self-optimization module synchronously distributes the updated parameters to each functional module, completing a closed-loop iteration of the entire system and realizing continuous self-optimization of the system.

[0027] Example 1: Image Input Scenarios for Consumer Mobile Phones This embodiment uses a low-resolution, single-target image captured by a consumer-grade mobile phone as input to illustrate the complete workflow of the system of the present invention.

[0028] The standard device color distribution described in this invention refers to the mean distribution of the RGB three-channel color histograms of more than 100 natural scene images pre-collected from an IEC 61966-2-1 standard sRGB reference display device under D65 standard light source (color temperature 6500K) illumination conditions. For remote sensing satellite multispectral images, the standard color distribution uses the atmospherically corrected radiance reflectance distribution of the Landsat-8 satellite Band2, Band3, and Band4 bands as a reference. Both sets of standard color distributions are pre-stored in the system storage module in the form of a 256×3-dimensional lookup table and are loaded into memory once during system initialization for later use.

[0029] The multi-device image input module receives images input from mobile phones, detects that the image belongs to the sRGB color space, and transmits the image data and device type information to the optimal transmission color alignment preprocessing module.

[0030] After receiving the sRGB color space identifier, the optimal transmission color alignment preprocessing module extracts the RGB three-channel color distribution histogram of the image. It uses the sRGB color distribution as the source distribution and the system's built-in standard device color distribution as the target distribution. The standard device color distribution refers to the mean distribution of the RGB three-channel color histograms of multiple images pre-collected from an IEC 61966-2-1 standard sRGB reference display device under a D65 standard light source (color temperature 6500K). This histogram is pre-stored in the system storage module in the form of a lookup table and loaded during system initialization. Based on the aforementioned cost matrix formula, an optimal transmission cost matrix is ​​constructed, and the Sinkhorn-Knopp algorithm is used for iterative solving to obtain the optimal transmission matrix. Based on this transmission matrix, the color distribution of the mobile phone image is mapped to the standard color distribution, completing color alignment. Simultaneously, the module activates downsampling preprocessing to appropriately compress the image resolution to reduce subsequent computational overhead, outputting a color-normalized preprocessed image.

[0031] The first Q-learning agent in the Q-learning convolution kernel dynamic configuration feature extraction module evaluates the complexity of the preprocessed image: the current image entropy is 0.3 and the target density is 0.1, both of which are below the threshold. The agent selects the first set of convolution kernel configurations from the Q table, namely 32 7×7 convolution kernels, to quickly extract image features in a coarse-grained mode, effectively reducing the amount of computation while ensuring feature quality, and outputs the image feature vector.

[0032] The reinforcement learning recognition decision module makes recognition decisions based on the image feature vectors and outputs the recognition results in this scenario. The inference latency reaches 10 milliseconds and the recognition accuracy reaches 98.2%.

[0033] The performance monitoring module collects data on recognition accuracy, computational latency, and memory usage during the recognition process. This data, along with input resolution, image entropy, and feature dimensions, forms a five-dimensional state vector, which is then passed to the adaptive adjustment module. The second Q-learning agent in the adaptive adjustment module selects a three-dimensional action based on the current state and the multi-objective Pareto reward function. The self-optimization module distributes the action parameters to each functional module to complete parameter fine-tuning. After several batches (1000 images per batch) of closed-loop iteration, the overall performance of the system in mobile phone image input scenarios continues to improve.

[0034] Example 2: Remote Sensing Satellite Image Input Scenario This embodiment uses high-resolution, multi-target images captured by remote sensing satellites as input to illustrate the workflow of the system of the present invention in highly complex scenarios.

[0035] The multi-device image input module receives satellite images, detects that the images belong to a multispectral color space, and transmits the image data and device type information to the optimal transmission color alignment preprocessing module.

[0036] After receiving the multispectral color space identifier, the optimal transmission color alignment preprocessing module extracts the color distribution histogram of the multispectral image, uses the multispectral distribution as the source distribution, and solves the optimal transmission matrix according to the cost matrix formula and the Sinkhorn-Knopp algorithm to complete the color alignment. At the same time, the module simultaneously activates contrast-limited histogram equalization processing to enhance the contrast of the remote sensing image to improve image quality and outputs a color-normalized preprocessed image.

[0037] In the Q-learning convolution kernel dynamic configuration feature extraction module, the first Q-learning agent evaluates the complexity of the preprocessed image: the current image entropy is 0.9 and the target density is 0.6, both of which are higher than the threshold. The agent selects the third set of convolution kernel configurations from the Q table, namely 128 1×1 convolution kernels, to accurately extract rich channel features in a fine-grained mode, ensuring the feature extraction accuracy in complex multi-target scenarios, and outputting the image feature vector.

[0038] The reinforcement learning recognition decision module selects a high-precision recognition strategy based on the image feature vector and outputs the recognition result in this scenario. The recognition accuracy reaches 99.1%, which meets the requirements for high-precision recognition of multiple targets in satellite remote sensing images.

[0039] System-wide self-optimization iterative process Whenever the system processes a cumulative total of 1000 images, a system-wide closed-loop iterative update is triggered. The performance monitoring module calculates the average recognition accuracy, average computation latency, and average memory usage of the 1000 images in that batch, forming a quantitative summary of the current system performance. The second Q-learning agent in the adaptive adjustment module substitutes the statistical data into the multi-objective reward function to calculate the current Pareto weight coefficient. , , The reward value is calculated and the Q-table is updated using a Q-learning strategy. The weight coefficients are adjusted according to the Pareto optimality principle to find a better balance between recognition accuracy, computational latency, and memory usage. The self-optimization module receives the updated 3D motion parameters and adjusts the preprocessing parameters accordingly. The data is sent to the optimal transmission color alignment preprocessing module to adjust the convolution kernel parameters. The learning rate is adjusted by sending the configuration to the Q-learning convolutional kernel dynamic configuration feature extraction module. The data is sent to the reinforcement learning recognition and decision-making module to complete a synchronous update of all system parameters, realizing the closed-loop self-optimization iteration of the system. As the number of iterations increases, the overall performance of the system continues to improve, with a cumulative performance improvement of 8.9% over a long period of time. Continuous performance self-upgrades can be achieved without manual intervention.

[0040] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. An image recognition system for artificial intelligence based on reinforcement learning, characterized in that: The system includes a multi-device image input module, an optimal transmission color alignment preprocessing module, a Q-learning convolution kernel dynamic configuration feature extraction module, an reinforcement learning recognition decision module, a performance monitoring module, an adaptive adjustment module, and a self-optimization module; The multi-device image input module is used to receive image data from multiple heterogeneous devices and to detect and identify the device type and color space type corresponding to the image data. The optimal transmission color alignment preprocessing module is connected to the multi-device image input module. It is used to extract the RGB three-channel color distribution histogram of the input image, construct the optimal transmission cost matrix between the source color distribution and the standard color distribution based on the color distribution histogram, solve the optimal transmission cost matrix using the Sinkhorn-Knopp algorithm to obtain the optimal transmission matrix, and map and align the color distribution of the input image to the standard device color distribution based on the optimal transmission matrix, and output the color normalized preprocessed image. The Q-learning convolutional kernel dynamic configuration feature extraction module, connected to the optimal transmission color alignment preprocessing module, includes a first Q-learning agent. The first Q-learning agent uses the entropy value of the current input image and the target density to form state features, and uses discrete combinations of the number and size of convolutional kernels to form an action space. The action space includes three sets of discrete convolutional kernel configurations: the first set has 32 convolutional kernels and a kernel size of 7×7, corresponding to a coarse-grained feature extraction mode; the second set has 64 convolutional kernels and a kernel size of 3×3, corresponding to a medium-grained feature extraction mode. The third configuration has 128 convolutional kernels and a kernel size of 1×1, corresponding to the fine-grained feature extraction mode; the first Q-learning agent adaptively selects a set of convolutional kernel configurations according to the current image state features to extract features from the color-normalized preprocessed image and outputs an image feature vector. The reinforcement learning recognition decision module is connected to the Q-learning convolution kernel dynamic configuration feature extraction module, and is used to make recognition decisions based on the image feature vector and output the image recognition result; The performance monitoring module is connected to the reinforcement learning recognition decision module. It is used to collect the current recognition accuracy index, computation latency index and memory usage index of the system, and to combine the three indexes with the current input resolution, image entropy value and feature dimension to form a five-dimensional state vector, which is used as the state input of the adaptive adjustment module. The adaptive adjustment module, connected to the performance monitoring module, includes a second Q-learning agent. The second Q-learning agent takes the five-dimensional state vector as input and a three-dimensional action space composed of preprocessing parameter adjustments, convolution kernel parameter adjustments, and learning rate adjustments as output. It calculates the reward value of the current action and updates the action policy based on a multi-objective reward function. The multi-objective reward function is the sum of three terms: a positive weighted value for recognition accuracy, a negative weighted value for computation latency, and a negative weighted value for memory usage. The weight coefficients corresponding to these three terms are adaptively updated by the second Q-learning agent during iteration according to the Pareto optimality principle. The self-optimization module is connected to the adaptive adjustment module, the optimal transmission color alignment preprocessing module, the Q-learning convolutional kernel dynamic configuration feature extraction module, and the reinforcement learning recognition decision module, respectively. It is used to distribute the three-dimensional action parameters output by the adaptive adjustment module to the optimal transmission color alignment preprocessing module, the Q-learning convolutional kernel dynamic configuration feature extraction module, and the reinforcement learning recognition decision module to complete the multi-objective closed-loop iterative update of the parameters of the entire system. The performance monitoring module, the adaptive adjustment module, and the self-optimization module together constitute the multi-objective Q-learning closed-loop framework of the whole system, realizing continuous self-optimization of the optimal transmission color alignment preprocessing module, the Q-learning convolution kernel dynamic configuration feature extraction module, and the reinforcement learning recognition decision module.

2. The image recognition system for artificial intelligence based on reinforcement learning according to claim 1, characterized in that: The elements of the optimal transmission cost matrix are determined according to the following formula: ； in, The first in the source color distribution The color value and the target color distribution in the _ ... The cost of transmitting individual color values The first in the source color distribution These are color values, specifically the color values ​​extracted from the image input to the device. For the target color distribution, the first Each color value corresponds to a color value in a standard device image. for and The cosine similarity between two color values ​​indicates that their hues are more similar, resulting in lower transmission costs. The first in the source color distribution The positional weight corresponding to each color value For the target color distribution, the first The positional weight corresponding to each color value This is a spatial regularization coefficient used to adjust the relative weight of color similarity and spatial distance in transmission cost calculations. The system pre-calibrates based on the device type.

3. The image recognition system for artificial intelligence based on reinforcement learning according to claim 1, characterized in that: The multi-objective reward function is determined according to the following formula: ； in, This represents the multi-objective reward value at the current moment. The normalized recognition accuracy of the current recognition result is set to a value of [value]. Real numbers within the interval; The computational delay for the current inference process, in milliseconds. The maximum acceptable inference latency threshold preset for the system, in milliseconds. This represents the memory usage of the current inference process, in megabytes. This represents the maximum available memory capacity of the system, in megabytes. and All The normalized dimensionless number within the interval, and Uniformity of dimensions ensures the mathematical validity of the weighted sum of the three terms. The Pareto weight coefficient corresponding to the recognition accuracy represents the proportion of positive contribution of recognition accuracy to the reward function; The Pareto weight coefficient corresponding to the delay is used to calculate the proportion of the negative penalty for the delay in the reward function. The Pareto weight coefficient corresponding to memory usage represents the proportion of negative penalty imposed on memory usage in the reward function; , , The second Q-learning agent adaptively adjusts itself in each closed-loop iteration based on the Pareto optimality principle.

4. The image recognition system for artificial intelligence based on reinforcement learning according to claim 3, characterized in that: The five dimensions of the five-dimensional state vector are, in order: the resolution of the current input image, the entropy value of the current input image, the number of feature dimensions of the current feature extraction layer, the recognition accuracy of the current system, and the computational delay of the current system.

5. The image recognition system for artificial intelligence based on reinforcement learning according to claim 3, characterized in that: The self-optimization module triggers a closed-loop iterative update when the system has processed a total of 1,000 images. After the update is triggered, the performance monitoring module calculates the average recognition accuracy, average computation latency, and average memory usage of the 1,000 images in that batch. The adaptive adjustment module adjusts the weight coefficients based on statistical results using the Pareto optimality principle. The self-optimization module synchronously issues parameter update instructions based on the updated weight coefficients, updating the preprocessing parameters of the optimal transmission color alignment preprocessing module, the convolution kernel parameters of the Q-learning convolution kernel dynamic configuration feature extraction module, and the learning rate parameters of the reinforcement learning recognition decision module, respectively.

6. The image recognition system for artificial intelligence based on reinforcement learning according to claim 1, characterized in that: The color space types recognized by the multi-device image input module include the sRGB color space of consumer-grade devices and the multispectral color space of remote sensing satellite devices; When the input image is detected to belong to the sRGB color space, the optimal transmission color alignment preprocessing module synchronously activates downsampling preprocessing to reduce the amount of computation. When the input image is detected to belong to a multispectral color space, the optimal transmission color alignment preprocessing module simultaneously activates contrast-limited histogram equalization to improve image quality.

7. The image recognition system for artificial intelligence based on reinforcement learning according to claim 1, characterized in that: When the entropy value of the current image is below 0.5 and the target density is below 0.3, the first Q-learning agent selects the first set of convolutional kernels to perform coarse-grained feature extraction; when the entropy value of the current image is between 0.5 and 0.7 and the target density is between 0.3 and 0.5, it selects the second set of convolutional kernels to perform medium-grained feature extraction; when the entropy value of the current image is above 0.7 and the target density is above 0.5, it selects the third set of convolutional kernels to perform fine-grained feature extraction.

Images