Image big data analysis method based on feedback enhancement

Abstract

The invention discloses an image big data analysis method based on feedback enhancement, and particularly relates to the technical field of big data analysis, and the method comprises the steps: carrying out the defuzzification of existing image data through an image defuzzification method, carrying out the step-by-step noise addition of an existing blurred image through a Markov chain, and obtaining the posterior probability of the image data; then, backward traversal is carried out along the Markov chain, and high-definition image data are generated through learning noise; designing a feedback enhanced generative model to guide the generation of high-simulation image data, and guiding the generation direction by enhanced learning to complete the generation of a controlled high-simulation image; obtaining a lightweight generative model by learning local parameters of the complex generative model; and designing a multi-layer attention mechanism analysis model, and completing image analysis by considering the fine features of the image data. The diversity of data is captured from different angles, and the understanding and generalization ability of the model to complex sequence tasks is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis, and in particular to an image big data analysis method based on feedback reinforcement. Background Art

[0002] With the popularization of the Internet, sensor technology and mobile devices, huge amounts of data are constantly generated and accumulated. Big data analysis reveals the patterns, trends and insights hidden in these massive data by mining, sorting and analyzing them. This data-driven approach helps companies make smarter decisions, optimize operational efficiency, and promote innovation and competitiveness. In the field of scientific research, big data analysis helps discover new knowledge, solve complex problems, and promote academic progress. In addition, big data analysis also plays an important role in the fields of medicine, finance, social networks, etc., helping to improve the quality of life and enhance the level of service. However, existing image data has problems such as data ambiguity and small amount of key data.

[0003] Image generation technology can synthesize image data with diversity, richness and complexity, providing more comprehensive and realistic data samples for big data analysis. This synthetic data helps to make up for the deficiencies or deviations that may exist in real data sets, and helps optimize the training and performance of machine learning models. In addition, image generation technology can also be used for data enhancement, expanding the scale of data sets, and improving the generalization ability of models. Through image generation, researchers and companies can better explore, understand and utilize huge data resources, accelerating innovation and development in the field of big data analysis.

[0004] Generative diffusion models show great potential in generating realistic samples, overcoming the posterior distribution alignment obstacle in variational autoencoders and alleviating adversarial target instability in generative adversarial networks. However, compared with GANs and VAEs, their sampling process inherently requires a more time-consuming iterative process.

[0005] Therefore, how to solve the problems faced by big data analysis, such as low data quality, low efficiency of model generation and poor performance of big data analysis models, has become a key issue that needs to be solved in the current field of big data analysis. Summary of the invention

[0006] To this end, the present invention provides an image big data analysis method based on feedback reinforcement to solve the problems raised in the background technology.

[0007] In order to achieve the above-mentioned object, the present invention provides the following technical solutions: a method for analyzing image big data based on feedback reinforcement, wherein the method defuzzifies existing image data by using an image defuzzification method, and gradually adds noise to the existing blurred image by using a Markov chain to obtain the posterior probability of the image data; then, traverses backward along the Markov chain to generate high-definition image data by learning noise; The design feedback reinforced generative model guides the generation of high-simulation image data, and the reinforcement learning guides the generation direction to complete the controlled high-simulation image generation; A lightweight generative model is obtained by learning the local parameters of a complex generative model; A multi-layer attention mechanism analysis model is designed to take into account the subtle features of image data to complete image analysis.

[0008] Preferably, the image deblurring method includes a forward denoising process and a reverse denoising inference process; assuming that the input data is x, the forward denoising process is expressed as: ; The value range of t is [1, 2, ..., T], is a hyperparameter of the Gaussian distribution variance; in the forward noise process, as the time step t increases, infinitely approaches pure noise; when T is large enough, Converges to standard Gaussian noise (0,I), is the posterior probability of the image data; The inverse denoising inference process is the probability distribution of model learning , to generate high-definition image data: ; in, Represents the data at time t The mean and variance of Indicates noise.

[0009] Preferably, the feedback reinforcement generative model includes a forward denoising process and a reverse denoising inference process: the forward denoising process is as described in the forward denoising process of the image deblurring method; The inverse denoising inference process is implemented through feedback reinforcement learning: the Q-learning update rule is as follows: ; Among them, Q(s,a) is the Q value of executing action a in state s; r is the immediate reward obtained; is the learning rate; is a discount factor used to balance the importance of current rewards and future rewards; s' is the new state after executing action a, and a' represents the optimal action selected under the new state s'; The exploration mechanism in the reverse denoising inference process uses the ε-greedy strategy to explore new sample spaces; and Select an action. In the feedback-reinforced generative model, action a indicates how to generate data for the next time step t, such as adding more noise, reducing noise, etc. The Q value is updated according to the action iteration process to select the optimal action and is updated through feedback through the reward signal; according to the feedback from the reinforcement learning agent, the parameters of the feedback-reinforced generative model are adjusted to obtain a controlled high-simulation image.

[0010] Preferably, the lightweight generation model is designed as follows: Complex generative models output soft labels: ; in, is the soft label received by the lightweight generative model; K is the temperature parameter that controls the degree of softening; is the output of a complex generative model; The comprehensive loss function contains the original data loss function and the cross entropy loss between the soft labels: ; in, is the total loss function; It is a data loss function based on hard labels; It is a cross entropy loss function based on soft labels; is a hyperparameter used to balance the two loss functions; The lightweight generative model is trained by minimizing a comprehensive loss function: ; in, are the parameters of the lightweight generative model.

[0011] Preferably, the multi-layer attention mechanism analysis model is designed as follows: The input image sequence first passes through three different linear transformation layers to obtain the Query, Key, and Value matrices respectively; then, these transformed vectors are divided into several heads, each with its own independent Query, Key, and Value matrices; for each head, a scaled dot product attention operation is performed, namely: ; Among them, Q, K, and V represent the Query, Key, and Value matrices respectively. represents the matrix transpose, Indicates the Key matrix dimension; Finally, the outputs of all heads are concatenated together and then fused through a linear layer to obtain the final attention output vector.

[0012] The present invention has the following advantages: Through the designed image defuzzification method and feedback-reinforced generative model, the existing image data is defuzzified, and targeted guidance is provided through fine-tuning algorithms to ensure purposeful exploration and controlled generation, generate highly simulated image data, and provide high-quality image data for big data analysis; the design of the lightweight generative model uses a smaller network to generate samples and minimizes the difference between the original samples and the generated samples, providing the optimal mapping for economical, controllable and accelerated generation. Compared with the prior art, the multi-layer attention mechanism analysis model designed by the present invention uses multiple attention heads to simultaneously focus on different aspects of the input sequence, thereby increasing the learning ability and expressiveness of the model, capturing the diversity of data from different angles, and enhancing the model's understanding and generalization capabilities for complex sequence tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The overall framework diagram provided for the present invention; Figure 2 A schematic diagram of the image deblurring method process provided by the present invention; Figure 3 A schematic diagram of the feedback-enhanced generative model and lightweight generative model process provided by the present invention; Figure 4 Schematic diagram of the multi-layer attention mechanism analysis model process provided by the present invention. DETAILED DESCRIPTION

[0014] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] like Figure 1 As shown, the present invention provides an image big data analysis method based on feedback reinforcement, which defuzzifies existing image data through an image defuzzification method, gradually adds noise to the existing blurred image through a Markov chain, and obtains the posterior probability of the image data; then, traverses backward along the Markov chain, and generates high-definition image data by learning noise; The design feedback reinforced generative model guides the generation of high-simulation image data, and the reinforcement learning guides the generation direction to complete the controlled high-simulation image generation; A lightweight generative model is obtained by learning the local parameters of a complex generative model; A multi-layer attention mechanism analysis model is designed to take into account the subtle features of image data to complete image analysis.

[0016] like Figure 2As shown, the image deblurring method includes a forward denoising process and a reverse denoising inference process; assuming that the input data is x, the forward denoising process is expressed as: ; The value range of t is [1, 2, ..., T], is a hyperparameter of the Gaussian distribution variance; in the forward noise process, as the time step t increases, infinitely approaches pure noise; when T is large enough, Converges to standard Gaussian noise (0,I), is the posterior probability of the image data; The inverse denoising inference process is the probability distribution of model learning , to generate high-definition image data: ; in, Represents the data at time t The mean and variance of Indicates noise.

[0017] like Figure 3 As shown, the feedback reinforcement generative model includes a forward denoising process and a reverse denoising inference process: the forward denoising process is as described in the forward denoising process of the image deblurring method; The inverse denoising inference process is implemented through feedback reinforcement learning: the Q-learning update rule is as follows: ; Among them, Q(s,a) is the Q value of executing action a in state s; r is the immediate reward obtained; is the learning rate; is a discount factor used to balance the importance of current rewards and future rewards; s' is the new state after executing action a, and a' represents the optimal action selected under the new state s'; The exploration mechanism in the reverse denoising inference process uses the ε-greedy strategy to explore new sample spaces; and Select an action. In the feedback-reinforced generative model, action a indicates how to generate data for the next time step t, such as adding more noise, reducing noise, etc. The Q value is updated according to the action iteration process to select the optimal action and is updated through feedback through the reward signal; according to the feedback from the reinforcement learning agent, the parameters of the feedback-reinforced generative model are adjusted to obtain a controlled high-simulation image.

[0018] Preferably, the lightweight generation model is designed as follows: Complex generative models output soft labels: ; in, is the soft label received by the lightweight generative model; K is the temperature parameter that controls the degree of softening; is the output of a complex generative model; The comprehensive loss function contains the original data loss function and the cross entropy loss between the soft labels: ; in, is the total loss function; It is a data loss function based on hard labels; It is a cross entropy loss function based on soft labels; is a hyperparameter used to balance the two loss functions; The lightweight generative model is trained by minimizing a comprehensive loss function: ; in, are the parameters of the lightweight generative model.

[0019] like Figure 4 As shown, the multi-layer attention mechanism analysis model is designed as follows: The input image sequence first passes through three different linear transformation layers to obtain the Query, Key, and Value matrices respectively; then, these transformed vectors are divided into several heads, each with its own independent Query, Key, and Value matrices; for each head, a scaled dot product attention operation is performed, namely: ; Among them, Q, K, and V represent the Query, Key, and Value matrices respectively. represents the matrix transpose, Indicates the Key matrix dimension; Finally, the outputs of all heads are concatenated together and then fused through a linear layer to obtain the final attention output vector.

[0020] According to the above embodiments, the performance of the method proposed in the present invention is verified, and the results are as follows:

[0021] It can be seen from the results that the lightweight generation model of the present invention provides high-quality data guarantee for subsequent image analysis. At the same time, the proposed multi-layer attention mechanism analysis model takes into account the subtle features of image data to complete image analysis, and has higher detection accuracy than the traditional CNN model under both data qualities.

[0022] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for image big data analysis based on feedback reinforcement, characterized in that: Methods The existing image data is defuzzified by the image defuzzification method, and the existing blurred image is gradually denoised by the Markov chain to obtain the posterior probability of the image data; then, the Markov chain is traversed backward to generate high-definition image data by learning the noise; The design feedback reinforced generative model guides the generation of high-simulation image data, and the reinforcement learning guides the generation direction to complete the controlled high-simulation image generation; A lightweight generative model is obtained by learning the local parameters of a complex generative model; A multi-layer attention mechanism analysis model is designed to take into account the subtle features of image data to complete image analysis.

2. The image big data analysis method based on feedback reinforcement according to claim 1 is characterized in that: The image deblurring method includes a forward denoising process and a reverse denoising inference process; assuming that the input data is x, the forward denoising process is expressed as: ； The value range of t is [1, 2, ..., T], is a hyperparameter of the Gaussian distribution variance; in the forward noise process, as the time step t increases, infinitely approaches pure noise; when T is large enough, Converges to standard Gaussian noise (0,I), is the posterior probability of the image data; The inverse denoising inference process is the probability distribution of model learning , to generate high-definition image data: ； in, Represents the data at time t The mean and variance of Indicates noise.

3. The image big data analysis method based on feedback reinforcement according to claim 2 is characterized in that: The feedback reinforcement generative model includes a forward denoising process and a reverse denoising inference process: the forward denoising process is as described in claim 2; The inverse denoising inference process is implemented through feedback reinforcement learning: the Q-learning update rule is as follows: ； Among them, Q(s,a) is the Q value of executing action a in state s; r is the immediate reward obtained; is the learning rate; is a discount factor used to balance the importance of current rewards and future rewards; s' is the new state after executing action a, and a' represents the optimal action selected under the new state s'; The exploration mechanism in the reverse denoising inference process uses the ε-greedy strategy to explore new sample spaces; and Select an action. In the feedback-reinforced generative model, action a indicates how to generate data for the next time step t. The Q value is updated according to the action iteration process to select the optimal action, and feedback is updated through the reward signal. According to the feedback from the reinforcement learning agent, the parameters of the feedback-reinforced generative model are adjusted to obtain a controlled high-simulation image.

4. The image big data analysis method based on feedback reinforcement according to claim 1 is characterized in that: The lightweight generation model is designed as follows: Complex generative models output soft labels: ； in, is the soft label received by the lightweight generative model; K is the temperature parameter that controls the degree of softening; is the output of a complex generative model; The comprehensive loss function contains the original data loss function and the cross entropy loss between the soft labels: ； in, is the total loss function; It is a data loss function based on hard labels; It is a cross entropy loss function based on soft labels; is a hyperparameter used to balance the two loss functions; The lightweight generative model is trained by minimizing a comprehensive loss function: ； in, are the parameters of the lightweight generative model.

5. The image big data analysis method based on feedback reinforcement according to claim 1, characterized in that: The multi-layer attention mechanism analysis model is designed as follows: The input image sequence first passes through three different linear transformation layers to obtain the Query, Key, and Value matrices respectively; then, these transformed vectors are divided into several heads, each with its own independent Query, Key, and Value matrices; for each head, a scaled dot product attention operation is performed, that is: ； Among them, Q, K, and V represent the Query, Key, and Value matrices respectively. represents the matrix transpose, Indicates the Key matrix dimension; Finally, the outputs of all heads are concatenated together and then fused through a linear layer to obtain the final attention output vector.

Images