Document Type : Original/Review Paper

Authors

• Matin Gholami ¹
• Seyyed Ali Zendehbad ¹
• Jamal Ghasemi ²

¹ Faculty of Engineering & Technology, University of Mazandaran, Babolsar, Iran.

² Faculty of Engineering & Technology, University of Mazandaran

Abstract

Deep neural networks are data-hungry and prone to overfitting, especially with limited training data and fine-grained visual variations. While region-mixing augmentations like CutMix serve as effective regularizers, their purely stochastic nature often creates a semantic gap—critical discriminative features are inadvertently masked or replaced, yielding suboptimal training signals. To address this limitation, the authors propose a novel family of adaptive augmentation strategies guided by Grad-CAM attention maps, injecting semantic awareness into the mixing procedure. The framework comprises three variants: Feature-Based CutMix and Saliency-Guided CutMix, which target salient regions for replacement (single peak and multiple high-saliency regions via fixed thresholding, respectively); and Region-Adaptive CutMix (RACutMix), which preserves salient regions by generating non-rectangular, saliency-informed masks using an adaptive percentile-based thresholding mechanism. RACutMix strategically augments only background or low-saliency regions with donor content, producing semantically coherent training examples that maintain visual and label consistency. Systematic evaluation on the Oxford-IIIT Pet benchmark with ResNet-50 compares five training paradigms under identical protocols with repeated trials and significance testing. RACutMix achieves 89.76% test accuracy, a substantial 10.23 percentage point improvement over standard CutMix (79.53%). It also demonstrates faster convergence, more stable loss dynamics, and consistent gains across precision, recall, and F1-score. Training overhead increases by approximately 1.5× time and 0.5 GB peak memory—a cost the authors quantify fully. This work bridges explainable AI and training regularization, demonstrating that internal model attention can be effectively recycled to guide learning toward more robust and discriminative feature representations, validating interpretability-guided augmentation as a powerful paradigm for fine-grained visual recognition tasks.

Keywords

• CutMix
• Data Augmentation
• Grad-CAM
• Generalization
• Image Classification

Main Subjects

• H.6.3.3. Pattern analysis