Image Feature Extraction Approach To Image Retrieval

With the continuous development of information technology and the popularity of smart camera phones,image retrieval has been applied to a wider applications such as copyright infringement detection,land-mark identification,information retrieval and image/video annotation.Extracting the image features with the feature extrac-tion approaches to represent the image is an important conponent in image retrieval tasks.Facing the difficulties and challenges from a variety of image affine transfor-mation and the large volume of data,how to extract and represent the image features correctly and effectively has become an important research content in image retrieval field.Therefore,it is of great academic and practical significance to study the image feature extraction approach to image retrieval.The main work in this paper is as follows:Firstly,we extensively research on the existing image feature extraction approaches and convolutional neural networks which are introduced into computer vision field recently.And then we introduce several detectors,descriptors and feature encoding methods.Secondly,in this paper,we analyse detectors based local feature extraction approach and CNN based global feature extraction approach.Based on analysis above,we propose a CNN based hybrid feature extraction approach.As demonstrated in ex-periments,hybrid feature capitalizes the geometric invariances of local feature and the distinctiveness of CNN feature.It receives the retrieval mean average precision of 82.5%on Holidays dataset,which improves the precesion on traditional manual design features by 12.3%.Satisfactory performance is observed from this hybrid feature representation.Thirdly,to reduce the influence caused by offset of rectangular patches,we propose a content based DCT-SIFT feature extraction approach.With the observation that the eigenvalues of frequency feature are greatly reduced with the increase of dimension,we only intercept part of the local features and splice them to the final image feature.In thus way,we improve the mean average precision of image retrieval result and reduce the storage costs.As demonstrated in the experiment,approach we proposed can reduce the influence caused by offset of rectangular patches effectively.