Computational Color Naming for Human-Machine Interaction

In this paper, we present two simple methods for automatic color naming, the first one learns the pixel-wise color name annotations from the color-chips and later method learns the image-wise color names from the real-world weakly labelled images. Color information is an important feature for many computer vision applications. Color name as a descriptor finds it’s applications in many real-world tasks such as Color Blind assistance, image retrieval and scene understanding. Color naming in images is a challenging problem due to shadows, view angles, illumination conditions and surface reflections. Manual labelling of color names for real-world images in applications like search engines, fashion parsing and tracking is a tedious task and time consuming. The proposed systems for color naming automates the process and avoids human labelling. These methods are based on superpixels in the CIELAB color space. We trained Random Forests classifier with color-chip dataset for pixel-wise color naming and for image-wise color naming it is trained on weakly color labelled image dataset. Both models are tested on real-world images for color name judgments. Experimental results shows that color names learned through these proposed systems have advantages in terms of implementation costs, speed of execution and can be used in real-time applications with lowcost hardware.