Adaptive Dual-Granularity Consensus Manifold Fusion for Unsupervised Image Segmentation

Abstract

In vision-based industrial quality inspection, uneven illumination, specular reflection, and coating inconsistencies often cause unsupervised segmentation methods to suffer from over-segmentation and erroneous region merging. To address this problem, this paper proposes an adaptive dual-granularity consensus manifold fusion method based on spectral consistency residual optimization. Dual granularity specifically refers to global view-level residual weighting and local topology-level residual correction. The global residual evaluates the relative structural consistency of color-spatial, local binary pattern (LBP), and Gabor graph views for consensus manifold construction, while the local residual refines unreliable edge connections to suppress cross-boundary leakage. Experiments on industrial painted metal surface images demonstrate that the proposed method outperforms six representative existing methods in suppressing under-segmentation across heterogeneous surfaces and over-segmentation within homogeneous regions.