Object Segmentation Based on the Integration of Adaptive K-means and GrabCut Algorithm

Image segmentation is a well-known topic in image processing, and it remains as a hotspot and focal point for image processing techniques. In this paper, we propose a hybrid segmentation method, combining an Adaptive K-Means clustering algorithm and a novel automatic GrabCut segmentation algorithm to improve the performance of the object segmentation from the scene image. 

The proposed method is divided into six steps: Firstly, the RGB image normalization step is introduced to eliminate light variation and remove bright and shaded regions. Secondly, RGB colour space is converted to L*a*b* colour space to maintain accurate colour balance. Thirdly, we propose a novel automatic GrabCut segmentation algorithm to eliminate user interaction and make the segmentation process faster. Fourthly, the Adaptive K-Means clustering algorithm and the proposed automatic GrabCut segmentation algorithm are combined to segment foreground objects from the background. 

Fifthly, the shape refinement step is used to eliminate occlusion, noise, and smear issues from the segmented image. Finally, morphological operations are carried out to enhance the segmentation performance. The performance of the hybrid segmentation method is assessed using the MSRA benchmark dataset.

Keywords: Segmentation, K-Means clustering, Image processing, Colour space conversion, Image normalization