<p>The 3D Gaussian Splatting (3D-GS) technique has significantly advanced 3D scene reconstruction and novel view synthesis(NVS). However, in large-scale scenes, inconsistent illumination often leads to appearance variations, posing persistent challenges for reconstruction. To address this issue, we propose PELR-GS, a perception-enhanced framework for large-scale 3D scene reconstruction. By dynamically predicting Gaussian colors and integrating perceptual optimization, our method improves view-dependent appearance modeling and cross-view perceptual consistency. Specifically, we design a Dynamic Color Spline Decoder (DCSD) based on Kernel-based Activation Networks (KAN) to replace conventional spherical harmonics, enabling more flexible and fine-grained modeling of complex nonlinear view-dependent appearance variations. Additionally, we incorporate a high-frequency enhancement loss during optimization to capture fine-grained image details and improve visual fidelity. We evaluate our approach on challenging large-scale datasets, including Rubble and Building from Mill19, and Residence and Sci-Art from UrbanScene3D. Experimental results demonstrate that our method outperforms existing NeRF-based and 3D-GS approaches, achieving state-of-the-art rendering quality.</p>