Thin structures such as power lines, pavement cracks, and lane markings often span only one to three pixels, which makes connectivity easy to lose during segmentation. Standard approaches that patch images or group pixels into superpixels can sever continuous lines or absorb them into the background before classification begins. Topology-aware losses can penalize connectivity breaks, but the authors of a new arXiv preprint argue that such losses cannot recover what the representation has already destroyed.

The proposed framework, SEMIR, replaces the pixel lattice with a parameterized graph minor. The contraction map is designed to preserve thin-structure connectivity, collapsing millions of pixels into tens or hundreds of boundary-aligned supernodes. This allows full-resolution inference without patching at scales up to 21 megapixels in the paper's experiments. A lightweight graph neural network classifies the reduced graph, and an exact map lifts the predictions back to pixel resolution.

Across three datasets—TTPLA for power lines, CrackSeg9k for pavement cracks, and SkyScapes Lane for aerial markings—the same pipeline with identical architecture, features, loss, and hyperparameters matches or exceeds domain-specific baselines on Dice, IoU, and Boundary F1. The authors also report at least a 4.6x reduction in mask fragmentation relative to SLIC at matched inference. Because this is a single preprint, these results are self-reported and have not yet been independently replicated.