Leaf Segmentation under Loosely Controlled Conditions

We propose a robust and accurate method for segmenting specular objects acquired under loosely controlled conditions. We focus here on leaves because leaf segmentation plays a crucial role for plant identification, and accurately capturing the local boundary structures is critical for the success of the recognition. Popular techniques are based on Expectation-Maximization and estimate the color distributions of the background and foreground pixels of the input image. As we show, such approaches suffer in presence of shadows and reflections thus leading to inaccurate detected shapes. Classification-based methods are more robust because they can exploit prior information, however they do not adapt to the specific capturing conditions for the input image. Methods with regularization terms are prone to smooth the segments boundaries, which is undesirable. In this paper, we show we can get the best of the EM-based and classification-based methods by first segmenting the pixels around the leaf boundary, and use them to initialize the color distributions of an EM optimization. We show that this simple approach results in a robust and accurate method.