An improvement on GrabCut interactive segmentation method based on input color spaces

Abstract

Interactive image segmentation methods that require an initial region(s) determination by a user and allow addition / removal of these regions are prefered in many areas especially in medical applications. In recent years, from these methods, the GrabCut interactive segmentation method has gained popularity due to its high success and ease of use. The studies on the improvement of the method are mostly concentrate on the energy function. In this work, the effect of using other color spaces as an input on the GrabCut method has been analyzed in detail, without changing the main steps of the algorithm. In addition to the original method that use RGB as an input color space, well-known color spaces HSV, HSV-conic form, YCbCr, CIELAB color spaces and also recently introduced promising color spaces CIECAM02, CIECAM16 and JzAzBz color spaces have been adapted to the method and tested on a public image data set VGG. Experimental results show that the CIE color spaces - especially CIECAM16-SCD-give consistent and superior performance (4.74% accuracy rate improvement) than the other methods on the GrabCut mehtod. The use of this color space on the GrabCut method is also proposed in this work.