Classification of camouflage-coated samples using a compact spectral Mueller-matrix descriptor

Reliable classification of camouflage-coated samples is relevant to optical inspection and material characterization, but samples with similar visible appearance may exhibit overlapping intensity responses and remain difficult to distinguish. Spectral Mueller matrices provide additional information on polarization transfer; however, directly using the full matrix spectra produces high-dimensional inputs, whereas individual polarimetric parameters describe only limited aspects of the coating response. This work proposes a 12-feature polarization-retention and coupling spectral-moment descriptor derived from white-reference-normalized spectral Mueller matrices (WPRC-SM12). The descriptor combines the white-reference-normalized intensity response, diagonal-retention mean coordinate, diagonal-retention anisotropy magnitude, and selected U-V coupling magnitude and summarizes the level, first-order trend, and second-order curvature of each component using low-order spectral moments. A dataset of 1620 measurements was acquired from nine classes of camouflage-coated samples over 510-800 nm under 12 incidence-reflection geometries. The proposed descriptor and five baseline descriptors were evaluated using nested grouped cross-validation, a fixed incidence-angle hold-out test, and leave-one-geometry-out validation. Under nested grouped cross-validation, WPRC-SM12 combined with LDA achieved an accuracy of 93.52% and a macro-F1 score of 93.50%. Across the five evaluated classifier pipelines, WPRC-SM12 also achieved the highest mean accuracy. For WPRC-SM12, linear SVM achieved an accuracy of 90.00% in the fixed 50° incidence-angle hold-out test. Under leave-one-geometry-out validation, LDA yielded an aggregate accuracy of 95.80% and a macro-F1 score of 95.78%. The remaining errors were concentrated in a limited number of sample classes and showed class-dependent sensitivity to broader changes in measurement geometry. These results demonstrate that WPRC-SM12 provides a compact and physically interpretable feature representation for the classification of camouflage-coated samples. The geometry hold-out results further suggest that representative angular coverage during model development can improve classification consistency.