Research And Implementation Of Image Search System Based On Deep Learning

The number of digital Images from the Internet and smart devices has been increasing exponentially as the advent of AI.Images can provide comfortable and convenient services to customers,while they also bring great challenges.It is a puzzle how to select the pictures which attract people from the massive image data and afterwards search for relevant and similar contents for users.As one of the important fields of computer vision research,image search can solve the problem of large-scale digital image search for similar content.There are 3 problems in Image Search when applying to research and application:(1)Image content features are difficult to describe discriminatively,which means the expression of image features by computer model algorithm is different from that of human visual perception.This maybe reduce the accuracy of image search.(2)It should be tackled that single metric function in the approximate neighbor search algorithm has a bias towards the image content feature similarity leading to the accuracy rate reduction.(3)At present,most of the application engines for image search are Text-based mothods,and there are many cases that image search results are irrelevant.As the problems mentioned above,this paper focuses on the research and analysis of the generation of efficient and semantic image content features and the rapid and accurate realization of large-scale image search using deep learning methods.Eventually,realize and develop an image search system for Internet shopping platform.The main research contents are as follows:(1)Aiming at the problem that image content features are difficult to accurately express,a joint model of image content semantic feature extraction and coding based on CNN-DeepAE is proposed and implemented.The model uses multi-feature fusion method to optimize the extraction of image content features,and at the same time design a multi-layer progressive progressive method to compactly encode the extracted features.Through the experiments on the CIFAR-10 dataset,the accuracy of image search is 4.86 percentage points higher than that of public paper research;(2)Aiming at the problem that the single metric function has the bias of image feature similarity results and the accuracy of approximate nearest neighbor search algorithm is reduced.Based on the Annoy algorithm,the construction method of multi-measurement overlay tree and the aggregation reordering algorithm are proposed,which increases the accuracy of the CIFAR-10 dataset by 1.8 percentage points.(3)Propose an image search system design scheme for the Internet shopping platform to optimize the results of image search at home and abroad.Based on the results of the model algorithm described above,this paper completed an image search system for the Internet shopping platform,which enables users to complete the corresponding business needs based on the API of the system and enhance the user experience.