Research On Some Key Technologies Of Image Retrieval Based On Convolutional Neural Networks

The frequent use of many platforms,such as Facebook,Twitter,MicroBog and WeChat,has led to the wide source and high output of image data.It's important to make the effective management and efficient retrieval of image information resources.How to accurately and efficiently retrieve and return the images that users need from a large-scale image database with rich visual and semantic information is the research hotspot and difficulty in the field of multimedia image retrieval.Image retrieval technology has been widely used in monitoring and security system,auxiliary medical diagnosis system,shopping consumption platform and so on.However,the existence of "semantic gap" and the low efficiency of image retrieval lead to the limited development of this technology.Therefore,there are some key technologies of convolutional neural networks,hash algorithm and object detection algorithm.Convolutional neural network,with its mechanism of simulating human brain,establishes a learning method of multi-layer neural network to analyze abstract high-level semantic features,effectively reducing or even eliminating the "semantic gap".The hash algorithm completes the mapping from high-dimensional to low-dimensional feature space,and completes the task of fast response to user retrieval by reducing the feature dimension and thus reducing the amount of computation.The object detection algorithm can accurately locate the object position and identify the object category,which effectively improves the accuracy of image retrieval.The main work of this paper is to improve the efficiency of image retrieval,which is researched from the following three aspects:(1)The research background,significance and current situation of image retrieval are summarized,and the research content and structure of the paper are introduced.Some key technologies of image retrieval are summarized.(2)Aiming at the problems of traditional image retrieval methods based on visual features,such as weak feature expression ability and high computational complexity of high-dimensional features,an image retrieval method based on Convolutional Deep Hashing network is proposed.By adding a Binary-like layer and designing a new loss function,we can get a compact hash feature with strong expression ability,and complete efficient large-scale image retrieval task in low dimension space.(3)Aiming at the problem that the SSD(Single Shot Multibox Detector)algorithm is not robust to small object detection and the speed of feature extraction is slow,a multi-object image retrieval method based on enhanced SSD-MobileNet(ESSD-MobileNet)is proposed.In this method,ESSD-MobileNet multi-object detection algorithm is used to locate the multi-object area of the image and extract the features of the multi-object area.Non maximum suppression method is introduced to screen the final multi-object area,and similarity measurement method is designed to get more reasonable multi-object image retrieval results.