Bio-Inspired Neural Radiance Fields for Enhanced Novel-View Synthesis in Complex UAV Scenes

Novel-view synthesis in complex UAV outdoor scenes remains challenging due to large spatial variations, weak or repetitive textures, distant structures, occlusions, and illumination changes. Inspired by the multi-scale receptive-field processing and selective feature-response mechanisms of early mammalian vision, this study proposes a bio-inspired Neural Radiance Field (NeRF) framework. In biological vision, receptive fields of different spatial scales contribute to the representation of both coarse scene structures and fine visual details. Following this principle, a multi-resolution hash-grid encoding strategy is employed to represent 3D coordinates across 14 spatial resolution levels, jointly capturing global scene layouts and local high-frequency details. Furthermore, inspired by the selective enhancement of informative visual features in early vision, a lightweight Directional Coordinate Attention (DCA) module is introduced. Through three independently parameterized gain branches, the module multiplicatively recalibrates intermediate feature channels to strengthen the representation of local structures and view-dependent appearance variations. A pre-normalized residual MLP and ray-wise density-variation and inter-level consistency regularization are further incorporated to improve optimization and rendering stability. Experiments on residential buildings, terraced fields, and mountainous terrain show that the proposed method consistently outperforms NeRF, Mip-NeRF, and Plenoxels in PSNR, SSIM, and LPIPS, achieving a maximum PSNR improvement of 2.27 dB over NeRF. These results demonstrate that introducing biologically inspired multi-scale representation and selective information processing into neural scene modeling can effectively improve novel-view synthesis quality in complex UAV environments.