The Research And Implementation Of A Web-based Clothes Image Retrieval System

The clothing image retrieval is a combination of traditional image retrieval technology and emerging electronic commerce. In recent years, with the bahavior of buying clothing online become increasingly popular, and the vigorous development of e-commerce apparel, clothing image retrieval technology is a research hotspot at present.The traditional image retrieval technology is very mature, but there are still many problems in the direct use of clothing e-commerce with large scale. It is necessary to avoid the huge workload of manual annotation of images, to ensure users getting search results before losing patience, and to guarantee enough retrieval accuracy to inspire user’s desire to use it again. These are basic requirements for image retrieval in clothing e-commerce.In this thesis, we develop an clothing image retrieval system. The main content include: The extraction of three CBIR features, The extraction of scale invariant interesting points, The clustering of feature points, Building unified descriptor based on the bag of words model.The extraction of three CBIR features is focused on color characteristic, shape characteristic and texture characteristic. Using RGB color statistical histogram, Canny edge detection operator and Tamura operator to extract the underlying fratures. Relevant examples are made to demonstrate this.In terms of detection of scale invariant feature points, first we filter the original images to construct scale space which is of translation invariant and scale invariant. Then find all extreme value points in the scale space, finally increase accuracy and reduce scope. The rest extreme value points are stable feature points. We use SURF algorithm to improve the efficiency considering the complexity and long operation time of SIFT algorithm.The clustering of feature points is to detect all scale invariant feature points on the data set of clothing images, then clustering SURF vectors of all feature points. We use K-means algorithm to cluster, which is a mature unsupervised learning algorithm in data mining. The disorder local feature points are normalized after clustering.Each image is expressed by the same frequency histogram under the bag of words model. Every frequency component is the clustering center got before.Finally we develop the Ocr clothing image retrieval system according to these algorithms. A experimental data set which having a number of categories, containing 10000 pieces of clothing images, is collected form Taobao online. We analyze and verify the performance of the system based on it. We check the system accuracy by analysing the experiment data under different K value and N value which is the retrieval return number. Compared with the traditional three CBIR characteristics, experiment data show that our algorithm get better precision and achieve better results.