<p>The depth of focus (DOF) in optical imaging is inherently limited, restricting the ability to capture sharp details across varying depth planes. Extended depth of focus (EDOF) techniques address this issue by fusing multiple images captured at different focal levels into a single all-in-focus image. However, traditional and handcrafted fusion methods often suffer from feature degradation, and limited adaptability to complex scenes. In this study, we propose a novel two-phase EDOF method that combines unsupervised deep feature extraction with spatial frequency-guided adaptive fusion. The first phase utilizes a lightweight auto-encoder architecture to extract hierarchical focus-aware features while preserving spatial resolution. In the second phase, a sharpness-aware fusion strategy based on local spatial frequency measurements integrates the extracted features to produce structurally coherent, all-in-focus images. Unlike prior fusion methods, our design explicitly separates feature learning from fusion, improving generalizability and efficiency. Extensive experiments on synthetic and real datasets demonstrate that our method outperforms state-of-the-art EDOF techniques, achieving higher PSNR and SSIM with lower MSE in synthetic settings, and superior perceptual quality in real datasets based on NIQE, BRISQUE, Entropy, and Perceptual Sharpness Index (PSI). The results confirm that the proposed method delivers a robust, adaptive, and computationally efficient solution for EDOF, with strong potential for applications in microscopy, remote sensing, and other precision imaging domains.</p>