Research On Shape-based Image Retrieval And Product Retrieval

In recent years, as an emerging e-business mode, online shopping is rapidly developed and widely accepted by people with its convenience, rapidity and low-costs. Many online shopping websites are coming out at the same time and a variety of products are shown with pictures on the website. In order to help people find the right product they want, content based image retrieval technology is used in e-business. But with the particular requirement of goods image retrieval, there are some technique problems to be solved such as the low matching accuracy and poor matching speed. In this paper, focusing on the characters of goods image, an exploratory research based on the shape feature of the image is given.Firstly, we introduce current researches, the characters, and the limitations of shape-based image retrieval. The most important and difficult thing is how to describe the shape efficiently, which decides the matching speed and accuracy. Traditional methods of shape matching require the completeness and continuity of a curve, and the solutions of discontinuous curve matching are rarely. But in real applications, there are often partly missing or discontinuous image contours. To solve this problem, we propose an approach for matching open curves based on the arc-tree representation and corner detection. It can efficiently describe the shape of a curve and have a good robust to the location, rotation, and scaling transform. Among them, the hierarchical representation simplified the feature of a curve and the corner detection of the maximum curvature gives an optimum starting point of matching. It reduces the dynamic programming matching computational complexity and raises the matching speed. Experiment result shows that our method provides a good way of fast matching with a good accuracy.Secondly, when there are many segments of an image contour, simply method of contour matching cannot meet the retrieval accuracy, and a more accurate method is needed. In this article, we use pyramid of histograms of orientation gradients (PHOG) descriptor to represent the shape feature. We combine it with the forward method, and get a good matching result.Finally, combining the rapidity of contour matching and accuracy of PHOG, we design a hierarchical retrieval system. In this system, coarser level matching procedure gives a general selection of the shape with a fast speed and finest level matching procedure gives a more accurate matching result. Experiment shows that our method gives a good result on product image retrieval and provides a solution to the CBIR application on e-business systems.