Research On Necktie Pattern Retrieve Based On Image Low-level Features

As one of the basic clothing in people’s daily life, necktie’s styles and patterns are variousand change rapidly. There are thousands of patterns stored in sample library of the necktie pro-duction companies, and designers create more and more new patterns every day. In traditionalmanner, necktie patterns need to be retrieved in the sample library by manual classification andnumbering. It cannot accurately describe the retrieval request to meet the needs of informationquery for different customers, and cost a lot of time and labor power. The image retrieval basedon computer vision can quickly find the same or similar to the existing patterns from sample li-brary according to samples or scanned images, overcome the shortcomings of traditional manualretrieval, and avoid repetitive design, thereby improving the efficiency of new products and en-hancing necktie patterns design capacity. According to the characteristics of scanned necktie im-ages, the basic research of necktie pattern retrieval based on low-level features of image was car-ried out in this paper.The main contents of this dissertation are as follows:(1) It elaborated the development history and present study of image retrieval technology,analyzed the characteristics of necktie images and the difficulties in necktie pattern retrieval, andprovided the idea to retrieve necktie pattern based on low-level features of image according tothe characteristics of scanned necktie images.(2) An review of the principle of retrieval method based on low-level image features wasgiven. It introduced several common low-level image features, similarity measure methods, andevaluation methods of image retrieval’s algorithm performance.(3) A smoothing filter with texture suppression algorithm was proposed according to thecharacteristics of scanned fabric image. It set the cumulative weights of both different colors’distance and neighborhood gradient pixel point by Gaussian function, and the linear filtering wasapplied to each pixel of the image, so that the pixel’s color value could be reserved inside smootharea while keeping the edge between the yarns of different colors effectively and preserving thenecktie image’s pattern. Compared with the conventional image smoothing algorithms, the algo-rithm proposed in this paper could obtain better smooth images while suppressing the yarn tex-tures to some extent and preserving the fabric pattern information. (4) A retrieval method combined with edge direction histogram and MSER features wasproposed. Firstly, the hybrid median filter and smoothing filter with texture suppression wereused to preprocess the scanned necktie image; secondly, the target image matched with thescanned images in library by edge direction histogram to obtain similarity score; thirdly, the tar-get image matched with the scanned images in library by the SURF features of maximum stableextremal regions (MSER) to obtain similarity score; finally, the similar necktie patterns was re-trieved by compositing the tow similarity scores. The experimental results show that the methodproposed in this paper can retrieve necktie pattern accurately.