Research On Image Retrieval Method Based On Visual Features And Semantic Features

Image retrieval is one of the research hotspots in the field of computer vision.With the advent of the era of big data,people's demand for image retrieval systems is becoming more and more urgent.This paper has carried out research on image retrieval based on the underlying features of the image and image retrieval based on the semantic features of the image.The main research contents are as follows:(1)Existing content-based image retrieval methods exist some drawbacks,such as low retrieval precision,unstable performance.To address these drawbacks,in this paper a content-based image retrieval method is presented based on multi-feature fusion of principal component,oriented-gradient and color histogram.The idea for the proposed method is: first,input image is grayed and flattened into a one-dimensional vector,and the first n principal components from the vector yielded by the PCA algorithm are extracted,in other word,input image is represented as a n×1 dimensional PCA feature vector.Secondly,to remedy color and orientation information missed by PCA,oriented-gradient and color histograms are used to extract orientation and color features respectively.Thirdly,extracted oriented-gradient and color histograms are merged with PCA features to generate the multi-feature representation of the input image.This paper confirms that the proposed multi-feature method can better represent an input image and can easily measure the similarity between images.The experiments are carried out and evaluated based on Corel-1000,Coil-20,Ghim-20 and so on,and two evaluation indicators such as Precision and Recall,and compared with the 4 state of art methods,experimental results show that the proposed method significantly overtakes the four methods.In addition,the proposed method also significantly outperforms the four methods in computing cost.(2)In order to solve the problem that semantic features of middle layers in deep neural network cannot fully represent an image semantic information,a hash image retrieval method based on deep shallow layer feature fusion is proposed,in which,firstly,a Merge Net structure is constructed to obtain deep and shallow fusion features.Secondly,a joint loss function,referred as to Joint loss,is defined,which is used to calculate contrast loss and cross-entropy loss for pair of input image and label image,and is carried out on the backpropagation training process of Merge Net.Experimental results show that,with combination of Merge Net and Joint loss,the proposed model can not only learn the existing image recognition degree but also optimally approach the semantic features of the image.Applied the proposed model for image retrieval test on datasets CIFAR-10,MNIST,and NUS-WIDE,it significantly outperform existed relevent image retrieval algorithms with respect to evaluation indicators MAP and PR.