Adaptive Feature and Loss Modulation for Symmetry-Breaking Small-Defect Detection in Automotive Glass

Detecting small, low-contrast defects on automotive glass is difficult because multiple factors act together: specular reflections, a wide range of defect scales, and repeated downsampling, which collectively obscure the very weak local deviations that distinguish a defect from an approximately regular background. This paper presents AFLM (Adaptive Feature–Loss Modulation), a framework that ties feature extraction to instance-dependent supervision. MDDC (multi-scale dynamic dilated convolution) adapts receptive fields so that the micro-defect detail survives; CSAF (channel–spatial adaptive filtering) regulates the channel and spatial responses where scales are fused; DWL (dynamic weighted loss) reweights each instance by its size, aspect ratio, and contextual complexity. On our GLASS-LW 2025 dataset, AFLM reaches 91.4% [email protected] and 88.4% F1-score at 7.3 ms per image, against 85.3% and 81.8% for the YOLOv8 baseline. It also keeps the highest [email protected] under geometric perturbation, with the smallest mean drop (Δ = 1.9, versus 4.8 for YOLOv8 and 4.2 for YOLOv11). The reflection symmetry score S_reflect separates defects from the clean background (area under the receiver operating characteristic curve, AUROC = 0.90), and, under real dual-sided inspection, AFLM attains a 95.4% top–bottom agreement with the lowest group-equivariance error, E(G)_real = 0.089. Together, these results give the symmetry-breaking interpretation an operational form and show that the framework’s symmetry awareness comes from the deployment geometry itself, not from training-time augmentation.