Where to Put Low-Rank Capacity? A Grouped-Split Study of LoRA for Industrial Defect Segmentation

Surface defects on magnetic tiles and steel strips are small, sparse, and visually heterogeneous. Full fine-tuning of a segmentation transformer can learn these patterns, but it is costly and data-hungry. We ask where low-rank adaptation should put its capacity, and how the result changes when whole filename groups are held out. We freeze a SegFormer-B0 encoder and adapt attention projections and MLP layers with LoRA. On the Magnetic Tile Defect dataset, using six filename groups as held-out test sets and three seeds, q/k/v/o plus MLP reaches 0.779 defect IoU, against 0.723 for attention-only q/k/v/o and 0.772 for full fine-tuning. The difference from Full is 0.7 points, with a 95% interval from − 2.0 to + 3.4 (p = 0.54); the study does not establish equivalence. Attention-only LoRA already supplies most of the adaptation gain, and the MLP adapters close the residual gap on this dataset. On NEU-Seg, a steel-surface dataset, the same configuration reaches 0.699, against 0.684 for attention-only LoRA and 0.715 for Full; the MLP gain is smaller and not significant. The configuration trains 0.543M parameters, of which 0.395M is the decoder head and 0.148M is adapter-specific. A BitFit baseline reaches 0.608, so the gain is not explained by a few bias terms. A bottleneck adapter is close to qkvo + MLP on both datasets. Random image splits compress the gap between attention-only LoRA and Full, while group-held-out splits expose a larger gap; we report both and treat the difference as split sensitivity rather than proof of leakage.