Ear Biometric Identity Verification Using Fused Classical Texture Descriptors and Support Vector Machines: A Robustness Evaluation on the Ami Ear Database

Ear images are a stable, contactless biometric trait, yet many studies report closed-set identification accuracy rather than verification error rates under degradation.This paper evaluates one-to-one ear verification with fused classical texture descriptors on the public AMI ear database (100 subjects, 7 views, 700 images).Local binary patterns (LBP), histograms of oriented gradients (HOG) and Gabor responses are extracted after CLAHE, weighted so each block contributes equally, and compared as pairs by an RBF support vector machine whose decision threshold is fixed on a validation partition.A subject-disjoint protocol (60/20/20 subjects) is run over seven scenarios (clean, rotation and scale, occlusion, low illumination, blur, noise, combined), four methods and five random seeds, giving 140 experiments, with accuracy, FAR, FRR, EER, AUC and paired Wilcoxon tests.On clean pairs the fused descriptor reached an EER of 14.07%, against 19.40% for HOG, 19.67% for Gabor and 21.60% for LBP, and was best for every query view.Under degradation the advantage disappeared: averaged over the seven scenarios, HOG gave the lowest EER (34.54%), followed by Gabor (37.60%), the fused descriptor (41.44%) and LBP (44.97%), and most methods approached chance under the combined scenario.An exploratory, milder combined scenario showed that removing the LBP block lowered the fused EER from 50.7% to 33.6%.Feature extraction took 2 to 3 ms per image on one CPU core.Fusion therefore helps controlled verification but not robustness, which motivates learned, augmentation-trained models.