Research On Sketch Based Commodity Images Retrieval

With the rapid development of Internet and multimedia technology,and social media are widespread,the number of digital images dramatic increase.How to retrieval useful information from huge amount of images become a hot spot for researchers.Traditional image retrieval methods are based on text information,it can't cover the increasing demand of users,and content-based image retrieval(CBIR)system arises at the historic moment.CBIR technology doesn't retrieval images according to image names,and no manual calibration is needed.It represent image content with image features.Typical CBIR system extract one kind of image feature,or a combination of different features while they build the image database,and features are deposit into a feature database.Usually an index structure is built at the same time,in order to improve efficiOency of retrieval.Then a similarity matching algorithm is adopted to compute the similarity of image feature and feature from database.The images are returned to user with the similarity sorting with user sketch.As smart phone and the tablet where user can input with a touch screen spread quickly,the way people interact with computer has change a lot.And touch has become the main way people input information.People can draw images in their mind with finger touching the screen.Because sketch match well with our cognitive of image contours,sketch-based image retrieval which is a branch of content-based image retrieval has drawn more attention recent years.This article go deep into the technology of image feature based image retrieval,and summarize existing sketch-based image retrieval algorithm and the research status at home and abroad.Same with content-based image retrieval,there are three main problems in sketch-based image retrieval:selection of image features,similarity measure between image features and index structure.We study the commonly used image features,and their performance in sketch-based image retrieval.Finally we select image gradient orientation field as the image feature of our system.The gradient orientation field is a field diffusion of the gradient orientation of image edges.In such way,the field reflect gradient orientation information on edge pixels,while on none edge pixels,it reflect the gradient orientation information of neighboring edge pixels.Before the computation of gradient field,we calculate the saliency map of images so that we can get the main contour information and diminish the influence of image background and textures.After we get the gradient orientation field,the field image is evenly divided into image cells,and for each image cell,a gradient field orientation histogram is calculated as the image feature descriptor.In order to achieve rapid image retrieval,we apply a bag of visual word model for image features.First all image features are clustered into so-called visual words.Then for each image,we can represent it with a combination of visual words,or a histogram of visual words where each bin is the frequency of the visual word.Finally,an inverted index is built on each visual word to achieve rapid retrieval.We download commodity images from taobao,and build a commodity image database.Experiments are conduct on the commodity database and a scene image database,prove the effectiveness of our algorithm.The experiment prove that our algorithm get a higher accuracy.