Transfer Learning Evaluation of DNNs for Image Classification
A study investigates how to select the best pre-trained model for image classification tasks using transfer learning. Eleven image processing models pre-trained on ImageNet were refined by adjusting output layers and network parameters, then applied to five target domain datasets. Accuracy, accuracy density, training time, and model size were measured in single-episode and ten-episode training sessions. The research aims to guide model selection for target domain requirements.
Key facts
- Transfer learning reuses knowledge from a source domain to enhance target domain learning.
- Eleven pre-trained models from ImageNet were evaluated.
- Five different target domain datasets were used.
- Metrics included accuracy, accuracy density, training time, and model size.
- Training sessions were conducted in one episode and ten episodes.
- Output layers and general network parameters were refined.
- The study focuses on image classification tasks.
- The technique saves training time, memory, and design effort.
Entities
Institutions
- arXiv
- ImageNet