Image Retrieval Method Based On Depth Learning Feature Extraction And Tree-hash Mixed Index

With the popularity of the Internet,especially the popularity of mobile Internet in recent years,more and more image resources appear in the network.The traditional text-based image retrieval technology which rely on manual tagging can not satisfy the growth of image in the internet,that's why the content-based image retrieval(CBIR)technology appear.For the CBIR technology,first we need to extract the feature of the picture and get a feature vector.For an image database,after the feature extraction,we get a feature vector database,and we build an index base on the feature vector database to improve the search process.About the feature extraction on picture,there are some traditional ways,such as Gabor filter?Color histogram?LBP?SIFT?HOG and so on.These traditional ways basically base on some image feature like gradient?texture or color which can some how express the feature of an image.Although these ways can get an good result,they can only extract the underlying visual feature,but not the semantic feature which can express the process of human recognition of images.The appearance of deep learning can solve this problem.Deep learning models such as Convolutional Neural Network(CNN)using human visual mechanism,to simulate the process of human recognition of images base on its layer-by-layer structure.The deep learning models can not only extract the high-level semantic features,but also a automatic learning process which is convenient to learn in massive data.After the feature extraction process we need to build an index base on the features to improve the retrieval process.There are mainly two kinds of index methods in CBIR,the tree based and hash based methods.Since the retrieval time efficiency of tree based methods are too low,the hash based methods are more popular.For the hash based methods when the retrieval process return more results the retrieval accuracy decreased rapidly,so this paper propose a tree and hash mixed method to solve this problem.We first use the information purification methods such as GINI index and information gain to split the data to get several purer subsets,ant then build hash index base on these subsets respectively.The experiment result shows that the tree and hash mixed method is better and more stable than the hash based method.