Clothing silhouette recognition based on detection of key points
Objective The clothing silhouette serves as an important feature for distinguishing and describing garments,and it holds significant relevance in various aspects such as consumer guidance in purchasing,personalized recommendations and customization services,design and production optimization,as well as market trend analysis and research.Previous research generally relied on manually defined key areas and designed complex algorithms to extract key point dimensions,resulting in low efficiency in discrimination.In order to achieve accurate and rapid clothing silhouette classification,this research focuses on runway dresses and proposes a clothing silhouette classification algorithm based on key point detection.Method A convolutional neural network was used in this research to predict ten key points including shoulders,chest,waist,hips,and bottom hem.These key points allow for the extraction of the clothing's silhouette from complex backgrounds,resulting in a silhouette image composed of lines.To extract simple clothing silhouette features,the DenseNet network was enhanced by incorporating the Sobel edge detection algorithm.The extracted features were then compared with the features extracted from a relative standard silhouette database using the cosine similarity algorithm.This approach ultimately enables the discrimination and classification of clothing silhouettes.Results The average error rates for each key point ranged from 0.046 to 0.205.The key points on the sides of the bottom hem had relatively larger average error rates of 0.205 and 0.204.This is mainly due to the deformation of the bottom hem caused by the model's walking movements,making it challenging to discern the key points of the dresses with trailing hemlines caused by stage lighting reflections.The shoulder key points had the lowest error rates,with values of 0.046 and 0.053.This is because the clothing texture stands out more compared to the surface of the human body,resulting in higher accuracy in key point localization.The waist key points had slightly higher error rates with values of 0.071 and 0.081.This is often due to the design of division lines at the waist in order to highlight body proportions,making the waist key points relatively easier to identify.In addition to quantitative analysis for evaluating the model,this study also performed key point detection on five representative images with different silhouettes.For the experiment,100 images of A-shaped silhouettes,100 images of X-shaped silhouettes,and 58 images of H-shaped silhouettes were selected.Compared to two commonly used convolutional neural networks,i.e.,VGG16 and ResNet50,the DenseNet-based silhouette classification showed a some advantage in accuracy.However,the average accuracy rate reached only 94.7％,which is not a significant improvement compared to other methods proposed in previous studies.When a Sobel layer was added to the DenseNet network,the edge features were sharpened,leading to improved accuracy in silhouette classification for various body shapes,under which circumstances,the average accuracy rate reached 95.9％.Conclusion In this paper,an intelligent classification method for clothing silhouette based on key point detection is proposed.Automatic extraction and classification of clothing silhouettes is achieved by key point detection and similarity algorithm.The experimental results show that the method leads to 95.9％classification accuracy in silhouette recognition,and the F1 score reaches 0.941.In order to improve the accuracy of convolutional neural network's extraction of edge features,the Sobel edge extraction algorithm is applied to the feature extraction process of DenseNet network.For comparison,whilst the convolutional neural network is able to learn the edge features in the image,the silhouette recognition method based on key point detection proposed in this paper is applicable to the silhouette recognition of various types of garments,which provides ideas and references for future silhouette classification research.
万方数据
维普
