Improving Leaf Segmentation by Creating the Automatic Marker Used In Marker-Controlled Watershed Segmentation

Abstract

Leaves segmentation in a state of nature is absolutely in need of a more complicated process. This is because the leaves that are captured outdoor may be included in a complicated background. The meaning of the complicated background is probably included with soil, residues and branches or overlapped/touched with other leaves that will complicate the segmentation process later. Most research related to leaf segmentation from complex background using watershed segmentation alone is inadequate. This is because, the technique is sometimes still segmenting leaves with imperfect conditions. To get the perfect leaf, post processing technique is needed to obtain the desired leaf. This will rise the time-consuming taken for processing and the technique for post processing should be developed to get the perfect leaf. Therefore, this research introduced the techniques which include the algorithm for automatic marker-controlled watershed transform without applying any post processing technique to obtain desired leaf. According to the previous study, over-segmentation will occurred if the watershed transform technique is directly applied to the gradient image. The problem occurs when there are irrelevant minima, other image irregularities and noise patches. Therefore, marker-controlled watershed transform is one of the approaches that can help decrease over-segmentation. A creation of marker is used to locate coarsely the objects and background. To improve the process of leaf segmentation using marker-controlled watershed transform, an improved algorithm of obtaining foreground marker was developed. The developed algorithm has an ability to create the foreground markers as needed for leaf segmentation. The foreground marker is determined automatically by combining techniques including morphological closing, morphological erosion and morphological reconstruction to the input images. This technique was applied to the gradient HSV images as input images while it is typically applied to binary images or gray scale images. The proposed algorithm will automatically create the foreground marker even though the shape of target leaf was irregular. From the experimental result, 74.1% of leaf images were successfully marked in order to segment individual leaf from complex background.