Efficient segmentation of leaves in semi-controlled conditions

We present a study on segmentation of leaf images restricted to semi-controlled conditions, in which leaves are photographed against a solid light-colored background. Such images can be used in practice for plant species identification, by analyzing the distinctive shapes of the leaves. We restrict our attention to segmentation in this semi-controlled condition, providing us with a more well-defined problem, which at the same time presents several challenges. The most important of these are: the variety of leaf shapes, inevitable presence of shadows and specularities, and the time constraints required by interactive species identification applications. We evaluate several popular segmentation algorithms on this task. Different datasets of leaf images are used, with manually segmented images serving as ground truth for quantitative comparisons. We observe that many of the methods are not immediately applicable: they are either too slow or would require that important modifications be introduced. We thus present extensions to our previously published segmentation method, which are able to improve its performance. The previous approach was based on pixel clustering in color space. Our extensions introduce a graph cut step and the use of a training set of manual segmentations in order to adjust important parameters of the method. The new method is fast enough for an interactive application, while producing state-of-the-art results.