Transfer Learning with Tiny YOLO-RV1 Deep Learning Network Model for Better Performance with Small Training Dataset

Abstract

In a desktop computer or laptop, training a deep learning network with large datasets is very difficult. The deep learning network will take several days for training the model on a higher end GPUs or TPUs even for a small-scale application. Hence, it decreases the interest of doing research in deep learning area with normal desktop computer or laptop. The objective of this paper is to make a YOLO model can be trained on a normal computer with a descent CPU power.

In this paper, we trained the standard tiny YOLOv3 model and the network complexity reduced version tinyYOLORV1deep learning network model for better performance with small training dataset. The scope of this work is to improve the performance of training with small dataset using transfer learning. After training the standard tiny YOLOv3 and the proposed tiny YOLO-RV1 models by transfer learning with few drone images (2 images only), their performances are compared with standard tiny YOLOv3 and tiny YOLORV1 models without transfer learning.

The arrived results proved that the possibility of training the deep learning network with a very few samples of images or any kind of data through transfer leaning achieves better accuracy. Further, our proposed Tiny YOLO-RV1 achieved an ideal performance in terms of accuracy even with the very small training set by using transfer learning.