Sketch-based Image Retrieval Using Improved HOG Algorithm

Today,with explosive growth of information and rapid development of multimedia technology and internet work,there is a huge demand on multimedia resource sharing among people.Therefore,an efficient information retrieval system to share a large quantity of images needs building.Existing social image searching technology mainly depends on text,images to provide users given query,which still fails to meet users' practical and general needs.Consequently,experts focus the study on image retrieval technology under the practical development of computer vision as well as big date.It is subjective and complicated to manually annotate images based on text through image retrieval.Meanwhile,with human-computer interaction and touch devices development,sketch retrieval has become a key part in content-based image retrieval.Social media users are able to get related images via image retrieval system by drafting through fingers or touch pen completely rather than manually annotation in keywords or illustration.The users can search out the relevant and desired images through the comprehensive feature description in the image retrieval system.The implementation process of sketch retrieval involves feature description,feature matching and related indexes.This paper focus on applying improved HOG(Histogram of Gradient)algorithm to sketch retrieval.In the first method,the proposed descriptor combines the Gaussian pyramid and local HOG feature descriptor.By using Gaussian pyramid,the image is decomposed into multiple resolution images on which the points of interest are extracted.Then points based multiscale HOG features are obtained.The algorithm expresses multiscale HOG features into visual words and finally forms feature vectors relative to visual words.The images are ordered according to decreasing similarity which is calculated through using the distance between the feature vectors of sketches and images.In the second method,On the basis of the multi-scale HOG algorithm,we improve the algorithm proposed in the first method.Firstly,we deal sketches and images with Gradient domain processing to minimize their inconsistencies.After that,a selection of interested points is made and image descriptors based on the selected points can be formed.This method can suppress noise on retrieval result.The experiment shows the effectiveness of improved HOG algorithm in the freehand sketch retrieval.Application of density gradient field strengthens edge information of freehand sketches,riches sketch features and optimizes the retrieval result.