Optimizing ResNet-50 for Multiclass Classification: A Multi-Stage Learning Approach

In this study, we present a multistage learning pipeline that utilizes the ResNet-50 architecture as a static feature extractor for multiclass image classification problems. This methodology integrates transfer learning, data augmentation, and adaptive learning techniques to enhance generalization across unbalanced and diverse datasets. We evaluated our approach using HAM10000, CIFAR-10, and CIFAR-100 to indicate its impact in both the medical and natural image domains. In contrast to providing a new network architecture, our contribution highlights a realistic and reproducible training schedule that compares effectively with present soft models such as EfficientNet-V2 and MobileViT-v2. Experimental results validate that our pipeline provides strong classification performance with minimal CPU resources, underscoring its applicability to practical image classification applications.