Research On Part Model Based Fashion Image Retrieval Technology

With the rapid development of image retrieval technology and e-commercetechnology, in order to satisfy the consumers’ demand for fashion shopping, a large numberof fashion image retrieval systems have come into being. However, we have been verifiedby experiments, the content-based fashion images retrieval engines’ retrieval accuracy isstill has much room for improvement. At present, content-based fashion images retrievalengines’ retrieval precision is not enough, such as the Tao Tao search, iSimilar. So, wedesign and implement a new algorithm which can effectively improve the retrievalprecision.Different from the general image retrieval technology’s global comparative method,this paper presents a Part Model Based Fashion Retrieval algorithm (PMBFR). The basicidea of PMBFR is to combine the specifyed image parts into part models as the basis ofsearch, and use the part models instead of the original image for the similar imagesretrieval. We eliminate the interference area of the fashion images by Character RegionLocation technology, and using the highest value local area in the image as comparisonbasis. This is according to the fashion image’s characteristic, which is that global area lookssimilar, and regional area looks different. By this method, we solve the two difficulties offashion image retrieval algorithm: the background interference and the error matching ofclothing category. Our algorithm has a good performance on fashion image retrieval.There are three main work and contributions of this paper.(1) we creativelyintroduces the object detection algorithm and the idea of part model, put forward a kind ofPMBCR image retrieval solution, which can effectively improve the accuracy of fashionimage retrieval. The system has four modules, including Character Region Location,Object Composition, Similarity Retrieval Based on Local Area and Algorithm Acceleration. Users can do experiment by adding the algorithm on other image retrieval systems.(2)Achieved several classical algorithms for each module of the PMBFR algorithm flow, andcompared these algorithms by experiment and statistical analysis. Finally we found anoptimal algorithm combination of the PMBFR algorithm flow. This is the core work of thesubject. The algorithm flow of PMBFR uses the Eclipse integrated environment asdevelopment platform, with Java language to implement. The system contains about totalof15,000lines code.(3) In order to solve the problem of insufficient clothing categorymatching accuracy, we need to accurate locate database image parts which are most similarto the query image parts specified by users.Combined the idea of Character RegionLocating Algorithm, this paper proposes an image similarity scoring algorithms, which issimilarity scoring algorithm based on the response of multi-scale expansion. Finally, weanalyze and compare the algorithm through theory and experiment. The experimentalresults show that the algorithm has a good performance in similar regional location.Experimental results show that, compared to the existing fashion image retrievalengines’ performance, the retrieval results of the proposed method’s precision rate has animprovement of10%, under the same recall rate. This shows that our method has a goodperformance on fashion image retrieval.