Segmentation of plant leaf images in natural scenes by integrating graph diffusion compactness and global refinement

Plant recognition, driven by advancement in image processing and machine learning, is crucial for ecological studies and agricultural management. However, challenges arise in complex backgrounds, requiring robust algorithms to differentiate target plants from surrounding elements. Leaf image segmentation is pivotal in accurately isolating and analyzing individual leaf structure, facilitating precise species identification. In this paper, we propose an algorithm for leaf segmentation in complex backgrounds. Our approach is notable for its low computational complexity making it well-suited for environments with restricted computational resources. First, we delimit roughly leaf areas in order to define the foreground template. Second, we rank the similarity of remaining image regions according to the foreground template through a graph based saliency process. Finally, the obtained contrast map is refined by using random forests to ensure optimal leaf/background separation. Experiments demonstrate that our algorithm outperforms several competing segmentation methods.