An image matching algorithm based on tensor subspace dimensionality reduction was proposed to the questions of easily losing relationships between pixels and intensively computational problems using traditional vector subspace methods. The algorithm extracts tensor subspace by employing two-sided projection transformation in edge images, reducing dimension of feature space and preserving the relationships between edge pixels. The algorithm measured the similarity between template and real-time image by calculating the correlation of dilated binary images. Experimental results on the standard face database and real IR images show that the new algorithm can improve the computational efficiency remarkably，and has a higher matching rate and matching precision than traditional vector subspace methods. The proposed algorithm can also be applied in cluttering and partially occluded circumstances.