Feature extraction algorithm based on dual-scale decomposition and local binary descriptors for plant leaf recognition

Plant leaf recognition is very important and necessary to agricultural information and ecological protection. Unfortunately, the robustness and discriminability of the existing methods are insufficient. This paper describes a novel plant leaf recognition method. In order to extract distinctive features from plant leaf images and reduce the probability of disruption by occlusion, clutter, or noise, a novel feature extraction algorithm based on dual-scale decomposition and local binary descriptors is proposed. The dual-scale decomposition consists of two phases. In the first phase, a plant leaf image is decomposed into several subbands with an adaptive lifting wavelet scheme. In the second phase, each subband is filtered using a group of variable-scale Gaussian filters. Local binary descriptors are extracted from the filtered subbands to capture both shape and texture characteristics, and then the histograms of the local binary descriptors at different scales and different subbands are determined and regarded as features. In order to improve the robustness and discriminability of plant leaf recognition further, a fuzzy k-nearest neighbors' classifier is introduced for matching. Experimental results show that the proposed approach yields a better performance in terms of the classification accuracies compared with the state of the art methods. It is also shown that this method is relatively robust to noise, occlusion and smoothing.