<p>This paper presents a novel illumination-invariant feature representation approach to address the challenge of undersampled face recognition under varying lighting conditions. First, an illumination level classification technique based on Singular Value Decomposition (SVD) is proposed to estimate the lighting condition of the input image. Second, we construct the Logarithmic Edge Feature (LEF) by combining the Lambertian reflectance model with local neighboring structures in the face image, applied across multiple spatial scales. Guided by the estimated illumination level, we then generate a high-performance LEF and adaptively fuse the multi-scale features to obtain the Joint LEF (JLEF) representation. Furthermore, a nonlinear constraint function is employed to suppress irrelevant high-frequency interference, enhance discriminative facial edge features, and generate the Adaptive JLEF face (AJLEF-face). Finally, we evaluate the proposed method on the Extended Yale B, CMU PIE, AR, and our self-constructed Self-build Driver database (SDB). Experimental results demonstrate that both the JLEF-feature and AJLEF-face consistently outperform existing methods, including deep learning-based approaches, for undersampled face recognition under severe illumination variations.</p>