Design And Implementation Of The Content Based Clothing Retrieval Model

With the development of society towards intelligent age,online shopping has become the main way of consumption,the process of consumers searching for goods is generally implemented by the retrieval algorithm of the e-commerce platform.The traditional Text-Based Image Retrieval(TBIR)algorithm can no longer meet the increasingly comprehensive needs when people doing online shopping.Benefits from the flourishing of deep learning,Content-Based Image Retrieval(CBIR)has become mainstream.Recently,grasping and aligning the local area information of the input clothing image is a popular direction in this field.On the basis of summarizing the existing work,following contributions are made to the extraction and alignment of local information of clothing images:1.Part alignment network based on attention mechanism is proposed.This network uses multiple local branch parallel architectures,and one single branch is a local feature extractor based on the attention module.Although each branch has the same structure,it can be adaptively learned to obtain different attention maps.In addition,a cross-domain sample mining algorithm is proposed,which is based on the deep model trained in the source domain and can digging and labeling unlabeled samples in the target domain.2.Multi-granularity part alignment network is proposed.This method performs a partition operation on the feature map rather than the input image to extract local information,and the performance of is improved by a large margin.And through the combination of horizontal,vertical and annular partition methods to ensure the integrity of the key part information.In addition,The proposed multi-granularity partition strategy effectively enhances the model by merging local features of different partition scales.The proposed two model design methods are based on different angles,which can effectively capture the local information of the garment,and the performance of the two models is further improved after fusion.This shows the two network models have good complementarity.