Low- and High-level Methods for Tree Segmentation

The detection/recognition of trees (trunks and branches) from 2D images is a challenging task in image processing. The large variety of visual appearance, environmental conditions, and occlusion make it an ill-posed problem. In our work we overview different approaches to solve this problem including the performance analysis of a pixel-level clustering and a neural network based approach. Besides discussing low-level approaches we proceed from low-level (pixel-level) representations to a high-level model using graphs. Thus the problem is transformed to fitting graph structures (vertices and edges) to 2D images based on appearance, and on prior information about trees. We use Reversible Jump Markov Chain Monte Carlo optimization to solve the energy optimization problem corresponding to the maximization of the probability of the graph model created in a Marked Point Process. Besides the color information (which can be modeled by Gaussian mixtures or convolutional neural networks) other properties (such as width, connections, overlapping of vertices, and direction of branches) can be coded by different energy terms corresponding to probability. Our new approach has the advantage that it does not require significant training and can result in a high-level graph representation. We present our initial results in this article thus the recognition of overlapping branches and occlusion by other trees is not presented in this paper.