Attention-aware color theme extraction for fabric images. Issue 5 (March 2018)

• Record Type: Journal Article
• Authors: Liu, Shiguang; Jiang, Yaxi; Luo, Huarong
• Is Part Of: Textile research journal. Volume 88:Issue 5(2018)
• Page Start: 552
• Page End: 565
• Publication Date: 2018-03
• Subjects: color theme -- fabric color -- fabric image -- image editing; Textile industry -- Periodicals; Textile fabrics -- Periodicals; Textile research -- Periodicals; Textiles et tissus -- Industrie et commerce -- Périodiques; Textiles et tissus -- Périodiques; Textiles et tissus -- Recherche -- Périodiques; Electronic journals; 677.005
• DOI: 10.1177/0040517516685278
• Languages: English
• ISSNs: 0040-5175

Abstract

Color configuration plays an important role in art, design, and communication, which can influence the user's experiences, feelings, and psychological well-being. It is laborious to manually select a color theme from scratch for handling large batches of images. Alternatively, it can inspire designers' creations and save their time as well by leveraging the color themes in existing art works (e.g. fabric, paintings). However, it is challenging to automatically extract perceptually plausible color themes from fabric images. This paper presents a new automatic framework for extracting color themes from fabric images. A saliency map is built to help recognize the visual attention regions of the input image. Since the saliency map separates the image into visual attention regions (foreground) and non-visual attention regions (background), we respectively compute the dominant colors of these two regions, and merge them to form the initial target color theme based on certain rules. The dominant colors are extracted, accounting for the characteristics of the fabric and hue distributions of fabric images so as to acquire visually plausible results. Our method can be used to transfer colors between two fabric images for fabric color design. We tested our method thoroughly with various fabric images (e.g. cotton, silk, and linen) with different texture patterns (e.g. plain and twill). Experiments show that our method is more efficient and can generate more visually plausible results than state-of-the-art algorithms.