WHEN IN-DOMAIN IMAGE RECOGNITION FAILS : A CONTROLLED DOMAIN-SHIFT STUDY OF HANDCRAFTED FEATURES FOR STEEL-SURFACE DEFECTS

Abstract

Industrial image-recognition models are often evaluated on randomly partitioned images generated under one texture and illumination process. This study quantifies how that protocol can misrepresent robustness. A reproducible 64×64 grayscale benchmark was procedurally generated with five balanced classes: clean surface, scratch, pits, inclusion, and scale. Training comprised 900 images; independent in-domain and shifted-domain tests contained 400 images each. The shifted set changed background orientation, noise level, and defect contrast. Four standardized linear-SVM pipelines were compared: downsampled intensities, histogram of oriented gradients (HOG), rotation-invariant local binary patterns (RI-LBP), and combined HOG+LBP. In-domain, the combined representation achieved the best macro-F1 of 0.577. Under texture shift, it collapsed to 0.101, while the simpler intensity baseline produced the highest shifted macro-F1 of only 0.301. RI-LBP retained 28.8% accuracy but macro-F1 was 0.238, indicating uneven class transfer. A severity stress test showed defect macro-F1 increasing from 0.245 for low-contrast anomalies to 0.358 for high-contrast anomalies. The negative result demonstrates that feature fusion can amplify source-texture dependence and that synthetic validation must deliberately change the generating process. These values are diagnostic of the controlled benchmark, not estimates for factory inspection; real NEU or MVTec validation is required.