Supervised global-locality preserving projection for plant leaf recognition

Abstract

Plant leaf-based species recognition is still a challenge due to the large intra-class differences and the inter-class similarity of nature plant leaves. A new manifold learning method namely supervised global-locality preserving projection (SGLP) is proposed for plant leaf recognition, including three stages. First, construct the local weighted inter-class and intra-class scatter matrices by local information and class information of the training samples, and then construct a global weighted inter-scatter matrix, finally design a multi-objection optimal solution to enhance the compactness of the intra-class points and inter-class-manifold separability. Compared with the classical manifold-based plant recognition methods, global weighted inter-scatter matrix is constructed to enlarge the distance between different classes of the data and then effectively reveal the intrinsic manifold structure for classification. Experiments on the ICL and Swedish leaf datasets validate that the proposed SGLP algorithm obtains higher classification results than other state-of-the-art methods.