In the field of high dimensional computer vision problems, the feature redundancy, high computation complexity and multidimensional structural loss are responsible for low recognition rates. This study introduces a Tensor Subspace Learning Framework (TSLF) which combines Higher-Order Singular Value Decomposition (HOSVD) and Tucker decomposition of tensors, low rank tensor optimization and Alternating Least Squares (ALS) to efficiently learn multilinear features and reduce the dimensionality. The framework is evaluated using classification accuracy, classification precision, classification recall, F1 score, feature preservation rate, reconstruction error, computational time, and memory usage and compared with the classification based on PCA, vector based SVD and CNN feature extraction. Experimental results reveal that proposed method achieved 97.86% accuracy, 96.91% F1-Score, 98.04% feature preservation and 29.8% less computational time when compared to other methods with the reconstruction error 0.021. The key novelty here is the consistent and mathematically sound framework for robust high dimensional computer vision analysis that is provided by a unified tensor decomposition and optimization that is scalable.