Image Classification Research Based On Fusion Of Deep Reinforcement Learning And Relation Network

Image classification and feature extraction are one of the important research contents in the field of machine vision and artificial intelligence.They have broad application prospects in many fields such as intelligent transportation system,auxiliary medical diagnosis,security monitoring and aviation detection.Conventional shallow-level classification techniques require design to be easily distinguishable or recognizable,which undoubtedly increases the workload and the recognition accuracy is low.At the same time,with the continuous development of multimedia information technology,many information features are constantly filled with the entire image,resulting in too much redundant information and irrelevant features in the image classification and feature extraction process,increasing the burden of the entire operation process and the computational cost,and affecting the efficiency and Accuracy.At the same time,due to the lack of image data or imbalance in the distribution of small-sample images in some special fields,it is difficult to produce satisfactory classification results and accuracy using conventional deep learning techniques.In view of these technical problems arising from image classification and feature extraction,this paper conducts the following research after learning from deep learning and strengthening learning related content:In view of the introduction of unrelated noise and redundant information in the image feature extraction process,which leads to the problem of image classification efficiency and low accuracy,this paper adopts image cropping model based on deep reinforcement learning.Deep learning has perceptual power,while reinforcement learning has decision-making power,and the two are combined to form deep reinforcement learning to achieve image cropping.Image cropping is regarded as a Markov model of sequence decision process,which allows the agent to interact with the cropping environment in reinforcement learning.By constructing a new reward function to guide the agent to optimize the action,it can find the best cropped image on the original image.By quantitatively analyzing the image cropping precision and the time efficiency in the cutting process on the relevant open source database,the cropped output image is visually and qualitatively compared to verify the validity of the model,and applied to the image classification model.In order to improve the accuracy and efficiency of small-sample image classification,this paper uses image cropping model and fusion relation network to solve this problem.The input image is cropped to extract image saliency regions,filtering out unrelated noise and redundant features.Then,the relation network is used for classification operation,and the known category image and the unknown category image are extracted by the feature extraction unit in the relation network,and then the extracted features are combined and the combined features are fed into the relationship metric unit in the relation network.Feature comparison is performed to realize small sample image classification learning.Through the experimental results,the classification accuracy and quantitative comparison are used to highlight the efficiency of the fusion algorithm.