Toward Amodal Instance Segmentation of Overlapping Concrete Aggregates via Occlusion-Conditioned Residual Correction

Concrete aggregate characterization is important for quality control, particularly in estimating particle size and shape distributions. Although computer vision provides a faster alternative to conventional sieve analysis, overlapping particles can hide parts of their boundaries, causing visible-only segmentation to underestimate the complete particle geometry. This study proposes an occlusion-conditioned residual correction mechanism for amodal instance segmentation of overlapping concrete aggregates. Rather than predicting the complete mask from scratch, the method preserves the observed visible region and estimates only the missing geometry along its boundary, using the visible mask as an explicit physical constraint. Local boundary information is incorporated through a ring context to guide residual prediction. The proposed mechanism is implemented within a standard Mask R-CNN framework and evaluated on a synthetic aggregate dataset with controlled occlusion levels. Compared with the corresponding direct amodal baseline, the proposed method improves AP from 48.02% to 52.69%, while ablation results show an additional gain from the ring context. Qualitative results further indicate more faithful reconstruction of irregular contours, narrow gaps, and sharp corners with less excessive mask expansion. These findings support constrained residual reconstruction as an effective mechanism for amodal aggregate segmentation under occlusion.