Recognition algorithm for plant leaves based on adaptive supervised locally linear embedding

Locally linear embedding (LLE) algorithm has a distinct deficiency in practical application. It requires users to select the neighborhood parameter, k, which denotes the number of nearest neighbors. A new adaptive method is presented based on supervised LLE in this article. A similarity measure is formed by utilizing the Fisher projection distance, and then it is used as a threshold to select k. Different samples will produce different k adaptively according to the density of the data distribution. The method is applied to classify plant leaves. The experimental results show that the average classification rate of this new method is up to 92.4%, which is much better than the results from the traditional LLE and supervised LLE.