TIER-BASED OPTIMIZATION OF IMAGE TRAINING DATA FOR EFFICIENT CONVOLUTIONAL NEURAL NETWORK LEARNING

Convolutional Neural Networks (CNNs) are widely recognized for their strong performance in image classification and other computer vision applications.Despite their success, training CNNs remains computationally intensive, largely due to the assumption that every training sample contributes equally to the learning process.In real-world datasets, many samples are repetitive or provide limited learning value, resulting in unnecessary computational overhead and longer training durations.To address this challenge, this paper introduces a Tier-Based Optimization (TBO) framework that prioritizes training samples based on their contribution to model learning.The dataset is divided into three tiers-high, medium, and low priority-using a composite importance score derived from loss values, gradient norms, and sample hardness.Samples identified as high priority are emphasized during training through weighted loss functions, while medium-and low-priority samples are strategically included at intervals to maintain robustness and generalization.In addition to data prioritization, the proposed method incorporates structured CNN pruning to eliminate redundant filters and channels, further reducing computational complexity.Experimental evaluations on standard benchmark datasets, including CIFAR-10, CIFAR-100, and a reduced ImageNet subset, demonstrate that the proposed approach achieves a 30-50% reduction in training time while preserving classification accuracy within 0.5% of traditional CNN training methods.These findings highlight the effectiveness of combining data-centric sample prioritization with structured pruning for scalable and efficient deep learning.