Comparative study of leaf image recognition with a novel learning-based approach

Automatic plant identification via computer vision techniques has been greatly important for a number of professionals, such as environmental protectors, land managers, and foresters. In this paper, we conduct a comparative study on leaf image recognition and propose a novel learning-based leaf image recognition technique via sparse representation (or sparse coding) for automatic plant identification. In our learning-based method, in order to model leaf images, we learn an overcomplete dictionary for sparsely representing the training images of each leaf species. Each dictionary is learned using a set of descriptors extracted from the training images in such a way that each descriptor is represented by linear combination of a small number of dictionary atoms. Moreover, we also implement a general bag-of-words (BoW) model-based recognition system for leaf images, used for comparison. We experimentally compare the two approaches and show unique characteristics of our sparse coding-based framework. As a result, efficient leaf recognition can be achieved on public leaf image dataset based on the two evaluated methods, where the proposed sparse coding-based framework can perform better.