Joint distances by sparse representation and locality-constrained dictionary learning for robust leaf recognition

Abstract

Plant species recognition has been a difficult and important task in agriculture, where computer techniques like image processing and pattern recognition can commendably facilitate plant recognition based on leaf images. The locality-constrained models produced by sparse representation and dictionary learning are a few of the prevailing feature models for leaf image recognition. Previous studies demonstrated that sparsity in representation plays an important role in the recognition, while sparsity constraints are the keys to solve the dictionary learning problems. Many of them focused on improving the sparsity, which is hard, but using large atoms in dictionary learning for high accuracy consumed more training time. Actually, sparse representation and dictionary learning are both based on distance calculation, e.g., Euclidean distance, which is also an aspect possible to obtain an improvement. On the premise of unchanged sparsity, this paper proposed a novel distance based method fusing Sparse Representation and Locality-Constrained Dictionary Learning (SRLC-DL) for robust leaf recognition. Integrating the distances obtained by dictionary learning and naive sparse representation can generate robust and high performance leaf recognition. In the fusion of distances, the number of atoms was not necessarily large as conventional methods, and even using smaller atoms produced more promising recognition at times. Therefore, not only has the leaf recognition accuracy by sparse representation been advanced, but the recognition speed also remains fast enough. A series of experiments had been conducted on five benchmark leaf datasets, including Caltech Leaves, Leaf, Herbarium, Swedish Leaf and Flavia. The experimental results demonstrated that SRLC-DL produced a higher accuracy in leaf image recognition and outperformed many other state-of-the-art methods.