Image Content Based Clothing Retrieval And Matching

With the development of apparel e-commerce, the clothing data on the internet is growing rapidly. Clothing retrieval and matching become the indispensable key technology. Clothing retireval can help the clients to efficiently search out the apparels they want. The ordinary users who are clothing non-professionanls can find the right dress collocation scheme through intelligent clothing matching technology quickly. How to invent a convenient and accurate clothing retrieval and an intelligent clothing matching technology for ordinary users becomes a very meaningful work. Due to the cognitive difference, it is very difficult to unify attribute semantics for the apparels in a large variety of styles. However, we find that each garment image contains a wealth of information. Based on image content, we do some researches on clothing retrieval and matching. The main contribution of the dissertation can be noted as following:1. The garment image feature extraction technology and description method are introduced from color, texture, shape and local features. We also analyze the characteristics of each feature to provide the computer vision theory for clothing retrieval and matching technology. The HSV color space, Tamura, Gabor, SIFT, SURF, and the maximally stable extremal regions(MSERs) are introduced with an emphasis. And a novel clothing skeleton extraction and node optimization algorithm based on constrained delaunay triangulation is presented.2. A novel bundled feature similarity calculation algorithm based on local word frequency and SIFTs distance matrix is presented to match bundled features. Each bundled feature consists of the point features(SIFT) which are further quantified into local visual words in a maximally stable extremal region(MSER). The similarity between two bundled features is computed from two aspects:1) local woerd frequency. We firstly check the common visual words in two bundled features, then the frequency of each common visual word related to each bundled feature is calculated. The local word frequencies and their difference are used to measure the similarity between two bundled features;2) SIFTs distance matrix. SIFTs distance matrix which is constructed by the distances between every two SIFT features in a bundled feature, represents the intrinstic geometric structure of a bundled feature and has its merits of scale invariance and rotation invariance. Experimental results show that our algorithm is effective.3. A novel image similarity fusion calculation method based on multiple bundled features is presented to realize clothing retrieval. We weight the importance of bundled features by IF-IDF, the precision of SIFT quantification and local word frequency. Then we can measure the clothing image similarity by computing the weighted similarities between their bundled features which have been matched. The clothing retrieval system can be realized through clothing image similarity and inverted-file index. Experimental results based on the clothing image database show that our approach works well in the situations with large clothing deformation, part hidden and background exchange, etc.4. A novel apparel style vision space definition method based on color and texture feature is presented. The HSV-T fusion feature is constructed with HSV color space and Tamura texture, and we can get the HSV-T subspace feature by KPCA. The apparel style vision space is difined by clustering the HSV-T subspace features. Experimental results show that the style vision space based clothing classification is similar with human visual sense.5. A novel intelligent clothing matching technology based on apparel style vision space and collocation data set is presented. We define the clothing collocation pattern based on clothing functional category and style vision space to construct the clothing collocation set is established with TPO principle. The frequent clothing collocation itemsets can be gained by Apriori algorithm. The dissertation establishs a clothing matching system with clothing correlation from mining the frequent collocation item sets. Experimental results show that our approach can guide the ordinary users find the right dress collocation scheme quickly.