Quality-Driven Reinforcement Learning for Adaptive Loss Weighting in Survey-Vehicle Removal of Mobile Mapping Panoramas

Mobile mapping system (MMS) panoramas routinely capture the survey vehicle, which occludes a substantial portion of the road surface beneath the camera. Reconstructing this large missing region is challenging because of complex road textures, cast shadows, and illumination variations. Although GAN-based inpainting typically combines multiple loss terms to improve reconstruction quality, fixed loss weights cannot adapt to the evolving optimization requirements during training. This study proposes a quality-driven loss-balancing framework in which a Proximal Policy Optimization (PPO) controller dynamically adjusts auxiliary-loss weights. At each control step, the controller observes interval-level generator losses, the current loss weights, and training progress, and receives a reward derived from paired improvements in SSIM and LPIPS on a disjoint controller-feedback set. Because ground-truth training data are unavailable, training pairs are constructed by projecting the occluded region into a perspective viewport and retrieving the corresponding background from nearby panoramas using GPS and heading information. The framework is evaluated against Fixed-Weight training, SoftAdapt, and GradNorm on two real-world datasets, MMS-Urban and MMS-Suburban, using Pix2Pix and Big-LaMa backbones. Experimental results show that PPO-guided weight adaptation achieves the lowest FID across all four backbone–dataset configurations. Relative to Fixed-Weight training, it reduces LPIPS, MAE, and FID by 14.80%, 15.64%, and 20.65% on average, respectively, while improving SSIM by 4.08%, with less than 1% additional training time. Further comparisons with predefined monotonic scheduling and progress-only PPO demonstrate the backbone-dependent contribution of state-informed adaptive loss balancing, with clearer reconstruction improvements for Pix2Pix and limited additional benefits for Big-LaMa.