COMPARATIVE ANALYSIS OF BLUR AUGMENTATION STRATEGIES FOR HARMFUL CONTENT IMAGE CLASSIFICATION USING EFFICIENTNET-B0: A PRELIMINARY STUDY

Automated content moderation at scale requires accurate and computationally efficient image classifiers, yet practical deployments often face severe data scarcity. This study presents a systematic comparison of six preprocessing strategies – Gaussian blur, Motion blur, Bilateral blur, Anisotropic blur, no augmentation (baseline), and background removal – for binary ethical/non-ethical image classification using EfficientNet-B0 trained from scratch on a self-collected small dataset (n = 472). All models were trained under identical protocols using class-weighted binary cross-entropy. Classification performance was evaluated on held-out test sets (n = 71 for blur conditions), with 95% confidence intervals computed via the Wilson score method and pairwise significance assessed by McNemar’s test. Bilateral blur and Motion blur achieved the highest observed test accuracy of 100.00% (95% CI [94.9%, 100.0%]; AUC = 1.0000). Anisotropic blur achieved 97.18% (95% CI [90.3%, 99.2%]) and Gaussian blur 90.14% (95% CI [81.0%, 95.1%]). The difference between Bilateral and Gaussian blur was statistically significant (χ² = 5.14, p < 0.05). Models trained without augmentation (37.23%) and with background removal (30.00%) failed to generalise. The edge-preserving property of the bilateral filter is proposed as a mechanistic explanation, pending Grad-CAM validation. These findings provide preliminary empirical guidance for augmentation selection in resource-constrained content moderation systems.