Automatic Classification and Color Changing of Saree Components Using Deep Learning Techniques

Abstract

Sarees are integral to Indian culture and serve as daily attire for most women on the subcontinent. Despite their popularity, there exists a gap in research regarding the automatic segmentation of sarees and the independent color modification of distinct components. Existing methods rely on labor-intensive manual adjustments through commercial applications, impeding productivity and resulting in avoidable expenses. This paper presents a tool that smartly coordinates different deep-learning techniques to modify the color patterns found on different parts of a saree. MODNet is applied for background removal and custom-trained Mask R-CNN models are utilized to precisely segment the saree body and border. The subsequent application of a color-changing algorithm in the HSV color space facilitates independent color modification for the saree border and body. The methodology proposed in this paper can be extended to any kind of clothing such as shirts, trousers, kurtas, kimonos, etc. An accuracy of 93.01% was achieved for the saree border segmentation, and an accuracy of 89.23% was achieved for the saree body segmentation when tested on a set of 50 test images.

Keywords

• Semantic segmentation
• neural networks
• color spaces
• MODNet
• mask R-CNN
• faster R-CNN