Plant leaf image segmentation in natural scenes: a multi-layer graph queries propagation approach

Accurate leaf segmentation is crucial for optimizing plant recognition and enhancing leaf identification precision. However, leaf segmentation encounters challenges when working with images captured in natural scenes. These images often contain intricate backgrounds with soil artifacts, overlapping leaves, plant elements, shadows, and variations in lighting. To address these issues, we propose an approach for segmenting leaf images using a multi-layer graph-based propagation method. The process begins with spatial localization of the leaf, aiding in detecting the foreground template, describing the central area of the leaf. Subsequently, a multi-level decomposition of the image into homogeneous regions is accomplished to capture image details at different scales. We then construct a graph based on this structure, connecting each region to its neighbors with weighted edges based on shared areas or edges across different resolutions. This graph is used to rank regional similarities to the leaf by propagating ranking scores from the foreground template to the image boundaries. As a result, we obtain a saliency map, which is used to extract the leaf from its surroundings. Finally, the resulting binary mask is refined using random forests to achieve optimal separation between the leaf and the background. Experiments conducted on a widely used dataset demonstrate that our method outperforms several state-of-the-art segmentation methods.