LLE Dimensionality Reduction And Its Application In The Infrared And Low-light Image Recognition

Due to the expansion of high-dimensional data, it is urgent to take effective measures on data dimensionality reduction. Since most high-dimensional data is non-linear, it is difficult for the traditional linear algorithm to find the structural features and correlation among high-dimensional data. While the nonlinear algorithm, especially the locally linear embedding(LLE), has been widely used because of its low computational complexity, less tuning parameters required and good robustness. This article conducts in-depth research and does corresponding improvement on unsupervised and supervised LLE. The main work includes three parts and they are as follows:1. This article conducts in-depth research for the LLE based on the Mahalanobis distance and the LLE with the selection of the optimal parameter of the neighborhood. The simulation results show that the LLE based on the Mahalanobis distance does better in performance on dimensional reduction and recognition rates. The optimal parameters of the neighbors for the sample using the double evaluation standard, the experimental results gained from two different data sets show that our method obtains better results.2. In this paper, aimed at the problem of the unsatisfying classification by theLLE, a further study on the supervised locally linear embedding algorithm(SLLE) is conducted. The SLLE algorithm based on the Mahalanobis distance and the selection of the optimal parameter is proposed, it improves the effectiveness and recognition rate and reduces the computational complexity.3. A new SLLE based on the index distance is proposed to solve the problem that neighbor parameter value of sparse samples is sensitive.The improved SLLE can enlarge the value range of the neighboring parameter and it has better robustness.The experimental results obtained from classic algorithm face database and captured images show that the improved SLLE has stronger generalization ability and gains better dimensional reduction and improved recognition rate.