Object recognition on controlled multi-view benchmarks can produce deceptively high clean accuracy, making it difficult to distinguish a genuinely informative representation from a model that mainly exploits stable appearance. This study develops and evaluates an interpretable recognition pipeline in which geometric information remains explicit while texture, color, and classifier diversity provide complementary evidence. The proposed Normalized Invariant Geometry descriptor (NIG-SHAPE) combines log-transformed moment invariants, normalized Fourier contour magnitudes, radial occupancy measurements, and compact contour geometry statistics. The 45-dimensional shape block is fused with a 26-dimensional Local Binary Pattern histogram and 96 RGB/HSV histogram features, producing a standardized 167-dimensional representation. Classification is performed by a Hybrid Feature Stacking Ensemble that integrates Random Forest, Linear Support Vector Machine, K-Nearest Neighbour, and Histogram Gradient Boosting predictions through a Logistic Regression meta-learner. Experiments use the corrected 7,200-image COIL-100 pipeline with a stratified 80:20 split and deterministic stress augmentation for the final model. NIG-SHAPE improves shape-only accuracy from 0.6424 for Hu moments to 0.8861, while the full hybrid representation reaches the clean-set ceiling. The robustness-augmented ensemble obtains 1.0000 clean accuracy and retains 0.9972 under occlusion, 0.9979 under lighting shift, and 0.9993 under clutter. The results show that the main value of the framework lies not in a clean-score superiority claim, but in a transparent combination of stronger geometry, complementary cues, classifier diversity, and disturbance-aware evaluation.

 

Keywords— object recognition; shape descriptor; feature fusion; COIL-100; stacking ensemble; robustness