Research On Image Retrieval Method Based On Region Of Interest

Modern Internet and mobile devices has been widely popular, a variety of large number ofimages has been everywhere in the network, full of people’s work and life. Also, at this stage thedemand for people to obtain images has greatly increased, image retrieval has become the mainway to get image. And because the difficulties for large number of image and heavy artificialworkload, the current major image retrieval mode have transformed from traditional text-basedimage retrieval into content-based image retrieval, which search image by image. Content-basedimage retrieval mainly relies on low-level image features to measure the similarity, which cannot effectively search to a satisfactory image, not focus on the central element of the query image.Therefore, the study of image retrieval methods based on the region of interest for query imagebecomes one of the current hot topics.The image retrieval methods existing usually merge single or global features to realize theretrieval. Image retrieval method based on the features fusion and adapt to complementary ofregion of interest and global image. So, we combine with the current machine learning methodsin image processing to propose an image retrieval method based on region of interest, which canimprove precision rate and recall rate of image retrieval, reflecting the query intent.This paper focuses on the region of interest for image retrieval, and it consists of thelow-level and high-level image feature extraction, similarity calculation method, image retrievalframework etc. The main innovative works of this paper can be summarized as follows:(1) A method of image retrieval based on the feature fusion of region of interest wasproposed to realize the semantic correlation of images content. First, the regions of interest weredivided and the integrated underlying characteristics of image were extracted. Second, thecharacteristics were used as training data to classify the images by semi-supervised learning, andthen the mapping between images and categories of semantic was established. Finally, thequadratic distance and the improved Canberra distance were respectively used for measuringlow-level features, and cluster centers of images in the feature space were updated iterativelythrough positive feedback. The experiments compared with other algorithms showed that theproposed image retrieval algorithm had higher accuracy and performed more effectively thantraditional algorithms.(2) In view of the visual dictionary in the requirements of image representation and retrieval,this paper proposes an image retrieval method based on the combination of multiple visualdictionaries and saliency weight, which achieves the representation of image features withsaliency and sparsity. Firstly, the image is divided into blocks and different kinds of underlyingfeatures of image blocks are extracted. Secondly, we use the image block features to learn themultiple visual dictionaries through non-negative sparse coding. The spatial information andsaliency is introduced into the sparse vectors for the image blocks by the method of saliencypooling, and saliency weight is introduced to form the sparse representation of the entire image.Finally, a proposed SDD distance is used for image retrieval. Experimental results comparedwith the method of single visual dictionary on the common image dataset Corel and Caltech demonstrate that the image retrieval method can improve the image retrieval accuracy.(3)This paper proposes approaches based on sparse coding and optimal componentsimilarity aggregation. Based on the idea that different parts of a face image have differentdiscriminative capabilities, and should contribute differently to the image recognition. Bypartitioning each image into un-overlapping and equal sized parts, and learning the sparseparameters and the reconstruction errors of each part for a query image, we construct a loss valuematrix. The reconstruction errors associated with each class is calculated by linear aggregationsbased on the matrices, and several weighted linear aggregation approaches have been proposedand discussed. Experimental results show the optimal weighted aggregation approachesoutperforms the classical SRC on most of the face images databases.