Using entropy based mean shift filter and modified watershed transform for grain segmentation

Abstract

Life science research aims to continuously improve the quality and standard of human life. One of the major challenges in this area is to maintain food safety and security. A number of image processing techniques have been used to investigate the quality of food products. In this paper, we propose a new algorithm to effectively segment connected grains so that each of them can be inspected in a later processing stage. One family of the existing segmentation methods is based on the idea of watersheding, and it has shown promising results in practice. However, due to the over-segmentation issue, this technique has experienced poor performance in various applications, such as inhomogeneous background and connected targets. To solve this problem, we present a combination of two classical techniques to handle this issue. In the first step, a mean shift filter is used to eliminate the inhomogeneous background, where entropy is used to be a converging criterion. Secondly, a color gradient algorithm is used in order to detect the most significant edges, and a marked watershed transform is applied to segment cluttered objects out of the previous processing stages. The proposed framework is capable of compromising among execution time, usability, efficiency and segmentation outcome in analyzing ring die pellets. The experimental results demonstrate that the proposed approach is effectiveness and robust.