Clothing Image Search System On Mobile Device

Traditionally,on-line shopping requires customers to enter the key word of wanted goods to search for corresponding commodity,which is inconvenient,as most of time image searching based on key-word is inaccurate.Furthermore,the user experience of this kind of application is terrible.The motivation of the work in this thesis is that in the mobile sight,taking the photo of commodity is convenient,and if we can search the corresponding commodity using that photo on on-line shop,the user experience will be highly improved.Based on the mentioned application background,this thesis will start the study on clothing image search on mobile platform.Our work included in this thesis can be summarized as follows: first of all,in consideration of the importance of feature extraction in image search and the resource-limited property of mobile phone,we compare many feature extractor used in visual word histogram construction in retrieval accuracy and energy consumption.Besides,we uses visual vocabulary tree to improve the quantization effectiveness of image feature.Based on the choose feature extractor,we also exam the sparse property of the visual word histogram of cloth image,which can heavily accelerate the effectiveness of network communication.In addition,to better fuse the different feature used in cloth image search,we propose the featurefused index based on product quantization,which is constructed by the Descartes product of different visual dictionary.As a result,it has better description ability,which have better performance on retrieval than traditional hierarchy inverted index for several feature.Furthermore,we also optimize the system based on and cosine distance algorithm,where the TF-IDF(Term Frequency-Inverse Document Frequency)are learned from text search engine and the cosine distance algorithm is based on the observation of the fact that cloth image has high degree of sparsity of visual word histogram.Experiments have been designed to evaluate the effectiveness of the designed algorithm in this thesis.Firstly,we evaluate the visual word construction algorithm based on exact cloth image set to indicate the better performance of visual word construction algorithm on energy consumption and search accuracy.Secondly,we design experiment for out proposed fused inverted index to indicate the effectiveness of fused-inverted index compared with hierarchical index used in traditional cloth image search system.Finally,we implement our system on physical architecture to indicates the effectiveness and applicability of our designed system on energy consumption,lower delay.