Research And Implementation Of Image Retrieval Technology Based On Multi-features Fusion

The image retrieval based on multi-features fusion can effectively overcome the shortcoming of deficient resolution of image description which based on single feature type.Therefore,people use more kinds of features to describe the image,and express image information by using complementary features which can effectively improve the retrieval accuracy.There are many kinds of image features and fusing methods,so it is of great significance to improve the accuracy of image retrieval by studying the combination of complementary features and efficient fusion.However,the local invariant feature of the image itself is stable and efficient,so combined the local features of image have great significance to describe the image content,but the Bag-of-features have a certain information loss,which can reduce the retrieval accuracy.While analyzing and improving the Bag-of-features,this article proposes a combined-features descriptor to weighting fusion of multi-features,which can help to design and realize the validity of the verification algorithm of the original image retrieval prototyping system.First,in order to deal with the information loss of the Bag-of-features,this article presents an algorithm of pictorial local feature representation-DIST Bag-of-features.The algorithm can extract the local invariant feature by constructing a multiple scale Dense SIFT descriptor.Feature quantization saves the distance information of each image between all visual vocabulary and local feature points.DIST Bag-of-features can use smaller dictionary to map and get the discerning representation vector of the local feature of the image.Using DIST Bag-of-features to describe the image can more accurately describe the local characteristics of the image and improve the retrieval accuracy.Secondly,point at the problem of insufficient resolution based on the single type,this article puts forward a combined-features descriptor of image which Using CNN to obtain the semantic information of images,weighted fusion CNN and other features to describe the image content comprehensively,and to further enhance the retrieval performance.This article combines the features of color and texture,DIST Bag-of-features and CNN features to describe images in order to obtain the information of image more comprehensively.In this experiment,the image databases of Corel1 K and Corel10 K were selected for the experiment,and the test results verified the effectiveness of this method.Finally,an image retrieval prototype system was designed and implemented by using the combined-features descriptor and the DIST Bag-of-features proposed in this article,which including offline multi-features extraction and feature library establishment,multi-features weighting factor setting,online query of images.It was verified by experiments that based on the feature algorithm of DIST Bag-of-features,the precision increased from 62.44% to 72.89% compared with the same type of algorithm based on the Bag-of-features.Based on the combined-features descriptor,the precision was increased from 82.60% to 84.67% compared with the joint feature descriptor.This result validly verifies the validity of the DIST Bag-of-features and the combined-features descriptor.