Extraction and classification of She nationality clothing via visual features

Abstract

The clothing pattern is an important feature of She nationality clothing, whose characteristics provide in-depth inference for digital computers to classify She nationality clothing automatically. In this paper, we propose a novel method to jointly reveal the texture characteristics and spatial distribution of the clothing pattern in She nationality clothing. Our approach employed visual features based on Scale Invariant Feature Transform and Gaussian blur to obtain robust and representative key points. We involved the Bag of Visual Words model and K-means clustering to construct a fundamental code word for texture representation. Also, spatial distribution was analyzed by frequencies of key points in spatial bins of the clothing’s design area. We then proposed a unified classification model applying the nearest neighbor method by utilizing both texture and spatial features. To demonstrate the efficiency of our method, we conducted 13 images of four types of typical She nationality clothing for training and testing. Our method achieved an average accuracy of 92.3% in cross-validation. The result revealed good consistency between the key points of texture and spatial features, and the different types of She nationality clothing characteristics. We further showed that our proposed algorithm could be used as inference for general clothing automatic classification and design.