The Research On Sketch Based Image Retrieval Algorithm

With widespread use of digital cameras, daily digital pictures greatly increase. Thus, theimage classification and retrieval has become an important application issue. But how tosearch in a large gallery of pictures is not yet solved effectively. In recent years, both fields ofdatabase system and computer vision, promote content-based image retrieval to become avery active area of research. Although content-based image retrieval technology has beenmade a great progress, some systems are also at the exploratory stage. The application has notyet fully reached the desired results, far away from a wide range of use. Therefore,content-based image retrieval technology has an important practical significance and greatvalue for research.Content based Image Retrieval Technology has two mainstream sub-fields, as Examplebased Image Retrieval and Sketch based Image Retrieval (SBIR). As an important techniqueto search a large image library, SBIR technique simplifies the query image as a sketch. Thiscan easily utilize users’ awareness for effective information extraction and facilitate retrievalto high-level semantic forward. SBIR can effectively overcome existing keyword-basedtechniques’ defect of requiring labeling, as well as make up for the need for suitable examplein example based techniques. However, users draw lines more sparse, the object with lineexpression does not have rich photometric characteristics. What’s more, objects drawn bydifferent people have different lines expressing styles, positions and sizes, even with differentnonlinear deformation. Therefore, the improvement of SBIR technology is challengingresearch issue.This paper proposes the improvement in both two categories of SBIR algorithm,respectively in global feature based method and local feature based method. Firstly, EdgeTangent Flow Field based multi-scale Structure tensor field retrieval algorithm is proposed.The algorithm calculates structure tensor feature on the Edge Tangent Flow Field, instead ofthe original calculation step directly on the gradient map of image. Beside, in the algorithmthe structure tensor feature is extracted in the form of multi-scale partition. The larger scale partition form has higher weight in the ultimate similarity metric. This is because that largerscale partition can avoid the influence of the gradient noise generated by the texture moreeffectively. The evaluating score of the proposed on the SBIR benchmark is higher that oforiginal structure tensor method. The experimental result proved that, the proposed methodcan effectively avoid the instability of the image gradient direction which indirectly describesthe significant edge of the image. As a result, the structure tensor can be calculated directly onthe remarkable strong edges, suppressing weak edge effects. And the multi-scale featureextraction method enhances the shift invariance to a certain extent. Secondly, Word of Interestbased SHoG Bag of Feature retrieval method is proposed. The algorithm uses a coherent linedrawing generation algorithm to improve the preprocessing step of image edge extraction.That has the advantage of that the generated lines can characterize the main lines of objects inthe image and be very close to user’s hand-drawn style realistically. The proposed methodutilizes the Word of Interest model in video retrieval algorithm to improve the existing SHoGBag of Feature algorithm. The modification effectively makes bag of feature associated withspatial relationships and eliminate defect of Canny operator stability. The experimental resultcan prove these improvements through the evaluating score on the SBIR benchmark.