Clothing Image Retrieval Based On Deep Learning For Extract Feature

With the rapid development of the Internet and e-commerce,the demand for people to quickly search for clothing in mass clothing images is growing.At present,the mainstream clothing retrieval method is text-based clothing retrieval,it requires a lot of manpower to mark the images in the image library,and the keywords of the user retrieve are too semantic,can not describe some clothing details,resulting in poor search results.Therefore,it is of great significance to find a method that can accurately retrieve the clothing image for the user.The content-based clothing image retrieval method has become an important research direction.In recent years,the deep learning method has been applied to the field of image retrieval due to its superior performance.Considering in the image retrieval based on the deep learning method,the image features largely determine the correct rate of retrieval.Therefore,this thesis mainly studies the extraction method of image features in clothing image retrieval based on deep learning,and completes the following work:A clothing image retrieval method based on the expression of attention area features is proposed.this thesis combines the extracted attention area with multi-label classification through a multi-label classification network based on attention area recursion,the extracted attention area features are used as local features of the clothing image retrieval.When extracting the attention area,this thesis proposes to constrain the network parameters through the loss function,to ensure that the attention area of the network extraction is moderate,and prevent the recurrent neural network from extracting the most significant and repeated attention areas each time.In the multi-label classification of clothing images,it is proposed to solve the problem of category imbalance by improving the input method.An adaptive classification loss function is proposed to solve the problem of imbalance between positive and negative samples in each category.Finally,a metric learning network is introduced in the multi-label classification network,and an input sample construction method based on the image label intersection ratio is proposed.Experiments show that this clothing image retrieval method can effectively improve the correct rate of clothing image retrieval,especially for clothing retrieval across scenes,and the retrieval performance is superior.Then,this thesis also proposes a clothing image retrieval method based on the expression of attribute space feature.Firstly,the clothing is detected according to the characteristics of the clothing image,and then we train the model of visual semantic joint embedding space,each type of attribute is modeled in the space,and the corresponding relationship between the attribute and the image position is obtained.Through this correspondence,the possibility of the attribute being expressed on the image point can be calculated,we can getting probability map of the attribute expression,and then the probability map is binarized,and the outer rectangular frame of the maximum connected region of the binary image is used as the local region corresponding to the attribute,and the region is mapped to the feature map extracted by the attribute classification network to obtain a fixed size local features,and finally the clothing image is retrieved in combination with the global feature and the local feature.Experiments show that the clothing retrieval method effectively improves the retrieval accuracy.