As a typical multilinear dimensionality reduction (DR) method, tensor locality preserving projection (TLPP) has been successfully applied in many practical problems. However, TLPP depends mainly on preserving its local neighbor graph which often suffers from the following issues: (1) the neighbor graph is constructed with the Euclidean distance which fails to consider the relationships among different coordinates for tensor data; (2) the affinity matrix only focuses on the local structure information of samples while ignoring the existing label information; (3) the projection matrices are nonorthogonal, thus it is difficult to preserve the local manifold structure. To address these problems, a multilinear DR algorithm called optimal neighbor graph-based orthogonal tensor locality preserving projection (OG-OTLPP) is proposed. In OG-OTLPP, an optimal neighbor graph is first built according to tensor distance. Then the affinity matrix of data is defined by utilizing both the label information and the intrinsic local geometric properties of the data. Finally, in order to improve the manifold preserving ability, an efficient and stable scheme is designed to iteratively learn the orthogonal projections. We evaluate the proposed algorithm by applying it to image recognition. The experimental results on five public image databases demonstrate the effectiveness of our algorithm.