<p>Existing learning-based ambient occlusion (AO) methods often struggle to balance high-quality detail preservation with real-time inference efficiency. To address this challenge, we propose FEAO, a lightweight feature-enhanced network for real-time AO reconstruction. The key idea is to improve feature representation from two complementary aspects: local geometric structure and broader contextual dependency. Specifically, we introduce a Gradient Extraction and Fusion (GEF) module, which employs learnable difference convolutions to extract geometry-aware gradient cues and enhance structural detail preservation. We further design a Lightweight Hybrid Attention Mechanism (LHAM) to improve contextual aggregation for AO prediction with low computational overhead. In addition, depthwise separable convolutions are adopted to reduce parameter redundancy and support efficient inference. Experimental results show that FEAO achieves the best image quality among the compared learning-based methods, reaching an SSIM of 0.903 while maintaining real-time inference performance (6.76 ms). These results demonstrate that FEAO provides a favorable quality-efficiency trade-off for practical AO reconstruction.</p>