Dual-Channel Scale-Adaptive Clutter Suppression with Confidence-Derived Gating for Microwave Inspection of Curved Radar-Absorbing Materials

Abstract

Microwave imaging is promising for non-contact inspection of radar-absorbing materials (RAMs); however, strong structured non-damage responses in reconstructed microwave images can resemble surface defects and trigger false detections, while hard masking may remove weak damage responses near ambiguous boundaries. To address these issues, this paper proposes dual-channel scale-adaptive clutter suppression with confidence-derived gating as an application-specific front end for surface-damage detection in curved RAM specimens. A normalized microwave-response map and a fixed morphology-enhanced map, both deterministically derived from the same measurement, are jointly encoded. The input-conditioned scale-adaptive pyramid-pooling module (SA-PPM) reweights multi-scale context branches, while confidence-derived adaptive-threshold gating (C-ATG) generates a bounded pixel-wise threshold from the retention confidence and its ambiguity proxy. Continuous-weight filtering (CW-FI) then produces continuous retention weights, enabling gradual attenuation instead of forced binary removal. Experiments were conducted on a paired visible-light/microwave RAM dataset. On the held-out test subset, the method achieved pixel-level discrimination between clutter and retained responses and improved detection of strip-like cracks and circular spalling under matched detector-side settings.