A Controlled Multi-dataset Evaluation of Custom CNNs, Pretrained Feature Extractors, and Transfer Learning

Convolutional neural networks (CNNs) are commonly trained using one of three paradigms: training from scratch, reuse of pretrained representations, or trans- fer learning via fine-tuning. While each strategy is widely adopted, their relative effectiveness depends strongly on dataset characteristics, computational con- straints, and deployment requirements. This paper presents a controlled and reproducible comparison of these three training paradigms across five real-world image classification datasets spanning infrastructure inspection, agricultural dis- ease recognition, and object-centric classification. All models are trained and evaluated under identical data splits, preprocessing pipelines, optimization set- tings, and evaluation metrics. Performance is reported using macro-averaged accuracy, precision, recall, and F1-score, alongside training time and model com- plexity. Rather than proposing new network architectures, this work emphasizes controlled evaluation and reproducibility to support evidence-based selection of CNN training strategies across diverse application domains.