Enhancement of individual leaf segmentation from complex background using watershed technique

In computer vision, difficulty in leaf segmentation from complex backgrounds still requires extensive research. This complex background is produced when the leaf image is captured in a natural environment. The uncontrolled sunlight conditions make it difficult to detect the green area of the picture. Even though there are many successful techniques used for leaf segmentation but the problem arise when we face with touching and overlapping leaves. Since both the target leaf and unwanted leaves have almost the same intensity values, the critical part is to segment the target leaf which usually touches/overlaps with other leaves which may create the confusion between the boundaries of adjacent leaves. Even though watershed technique is a powerful tool for separating touching or overlapping objects unfortunately, watershed transform itself for separating individual leaf from complex background leads to over/under-segmentation. The aims of this study is to eliminate non-green complex background, to enhance gradient of touching and overlapping edges (blur edges) between single leaf and non-target leaf and to improve the efficiency of watershed marker control for single leaf segmentation from complex background. In this study, three algorithms were introduced to assist in the segmentation of leaves from complex backgrounds. At first stage, Modified Excess Green Vegetation Index’ (MEGVI) was introduced to overcome the uncontrolled light problems in limiting non-green areas. Individual leaf segmentation process is done by the proses called marker-controlled watershed transformation. Watershed transforms actually have disadvantages where over/under-segmentation often happens. In order to overcome over/under-segmentation problem, the other two algorithms were introduced and could help to solve the problem. Since leaf captured outside are usually touching/overlapping with other leaves, the edges between the objects in the image are enhanced to aid thesegmentation process. The process of segmentation could be more effective if the accurate marker is created. By running these three algorithms, the desired individual leaves can be segmented perfectly without over/under-segmentation. The proposed technique gave 74.1% successful rate of leaf segmentation from complex background as compared to classical watershed segmentation of 0%, normal marker-controlled watershed segmentation, 13.8% and Xiaodong Tang watershed segmentation of 9.5%. From the experimental results, MEGVI is capable to eliminate non-green background in uneven illumination. While, the proposed gradient enhancement gave the best result in order to enhance the gradient image for touching and overlapping objects. A new algorithm to automatically obtain the foreground and background markers was proposed to improve the efficiency of the watershed marker control for single leaf segmentation from complex background. This approach is also benefits in creating foreground marker for irregular objects. The proposed technique also could reduced over segmentation/under-segmentation when applying watershed transform.