Abstract: As an effective nonlinear dimensionality reduction tool, the local tangent space alignment algorithm (LTSA) can obtain the global low-dimensional embedded coordinates of sampled data from a high-dimensional space. Introduced into multi-pose ear image recognition, LTSA has to improve to solve its problems in image recognition. An adaptive neighborhood selection strategy is proposed and a novel multi-pose ear recognition method based on this improved LTSA is present. Experimental results illustrate that it is an effective multi-pose image recognition method which can obtain better recognition rates than the traditional linear ones when the pose varies a lot