Research On T-shirt Attributes Recognition Based On Deep Learning

The continuous improvement of people’s aesthetic level and the rapid development of the Internet have prompted the fashion industry to contain huge market potential.The intelligent analysis of fashion clothing,from the fields of multimedia,pattern recognition,and computer vision has attracted widespread attentions from both academia and industry.Among them,the recognition of clothing attributes and key points is one of the basic issues for subsequent intelligent analysis tasks such as trend prediction and retrieval,so it owns important research significance.However,the clothing engineering knowledge system is highly specialized.At present,there is still improvement room in the research of this field based on deep learning algorithm,such as simple classification of single clothing attributes,and the accuracy needs to be improved.T-shirt is one of the most common types of clothing for people.This paper takes T-shirt intensive attribute instance segmentation and key point prediction as the goal,combined with the characteristics of the T-shirt design system and images,and focuses on the key issues of T-shirt clothing attribute recognition.We conduct the research from the following areas:(1)T-shirt intensive attribute data set creation.Combining the professional knowledge of clothing engineering,we summarize the attributes of T-shirts into four dimensions: collar,sleeve,hem,and fit,and design the corresponding sub-attributes for each dimension.The data set contains 20130 images.The annotation information not only includes the attribute classification information of the T-shirt,but also includes the location information corresponding to each attribute,so that the deep learning model can complete the instance segmentation task.(2)Diversified data preprocessing.We adopt four kinds of data enhancement methods:CLANE,IAASharpen,IAAAEmboss and Random Brightness Contrast.Through the above four data enhancement methods,we make model is not easy to overfit,and the model is more generalized.(3)Aiming at the feature extraction backbone network,an improved deep residual convolutional network T-Net is proposed.By adopting a delayed subsampling operation in the original residual convolutional network,the convolution kernel can fully extract the feature information of the feature map,thereby improving the detection ability of the model.(4)Improve the filtering prediction box algorithm in the original detection network.The classic non-maximum suppression(NMS)algorithm is improved to the soft non-maximum suppression(Soft-NMS)algorithm,thereby enhancing the detection performance of overlapping small targets.