Effective scale‐generation scheme based shape descriptor for fast leaf recognition and retrieval

Abstract

Plant species recognition using leaf images is a very challenging task in the field of pattern recognition. In this paper, an effective scale-generation scheme is proposed to extract leaf contour features. For each point on sampled leaf contour, the scale-generation scheme iteratively utilized trisections of leaf contour to find paired neighbour points under different scales. Then paired neighbour points under each scale were used to extract angle information to form multi-scale angle representation. Subsequently, Fast Fourier transform was applied on the multi-scale angle representation to make it compact and facilitate similarity measurement. City block metric was used to compute similarity between leaves for retrieval task. Both Support Vector Machine and 1-nearest neighbour were used as classifiers for recognition. Finally, the proposed descriptor was evaluated on four challenging leaf datasets, including Swedish, Flavia, MEW2012 and ImageCLEF 2012 datasets. The recognition accuracy of the proposed descriptor over Swedish and Flavia datasets reaches 97.03% and 94.03%, respectively. The mean average precision scores over Swedish, Flavia, MEW2012 and ImageCLEF 2012 datasets are 77.39%, 69.47%, 47.51% and 37.13%, respectively. Performance comparisons with both classical and state-of-the-art methods were made in terms of performance evaluation metrics. The results demonstrate that the proposed method has prominent performance.