Clothing Attribute Recognition Based on RCNN Framework Using L-Softmax Loss

Abstract

Due to the significant potential values in commercial and social applications, clothing image recognition has recently become a research hotspot, among which clothing attribute recognition is an important content. However, the large variations in the appearance and style of clothing and the image's complex forming conditions make the task challenging. Moreover, a generic treatment with deep convolutional neural networks cannot provide an ideal solution. Instead of using CNNs for classification, we proposed a novel approach based RCNN framework for the recognition task. Firstly, we apply the modified selective search algorithm to extract the region proposal. Then, the Inception-ResNet V1 model with L-Softmax is employed to represent images and identify their categories. After Soft-NMS, we use a simple neural network to correct the boundary of region box. To evaluate the performance of the framework, a dataset including about 100,000 shirt images was built. The experimental result show that our proposed framework achieved promising overall labeling rate, precision and recall of 87.77%, 73.59% and 83.84%. In addition, comparative experiments demonstrate the superiority of the proposed framework.