In order to keep the inherent higher order structure and correlation in the original data, reduce the influence of illumination in image recognition, and optimize the weight of the multi-scale feature, the method of image feature extracting based on weighted multi-scale tensor subspace is proposed to solve the problems. Firstly, multi-scale transform is used to characterize each place feature of the image, and uncertainty weighed is adopted on the role of each scale feature for image classification. And then a multi-scale tensor space is built using multiple linear principal component analysis and linear discriminant analysis algorithm to reduce the cost of processing, preserving the inherent structure and correlation of high-dimensional data. Finally, the extraction of the image features is completed. CAS-PEAL-R1 oriental face database is chosen for evaluation. The experimental results show that the algorithm performs better than some recent algorithms for image recognition with practical feasibility.