Transfer Learning on EfficientNet for Maritime Visible Image Classification
Mostafa Hamdy Salem, Yujian Li, Zhaoying Liu · 2022 7th International Conference on Signal and Image Processing (ICSIP) · 2022
The state-of-the-art in computer vision has significantly improved over the last few years as a result of the rapid development of deep learning techniques, the existence of large labeled datasets like ImageNet, and GPU acceleration developments. With the fast advancement of pre-trained models, transfer learning and fine-tuning are becoming more popular strategies for saving time and solving the problem of a shortage of data in image processing. For ship recognition, it is still a big challenge to obtain large labeled datasets like ImageNet. In this paper, we experiment with classifying the Marvel vessel dataset by reducing the training samples by half using transfer learning and fine-tuning strategies based on pre-trained EfficientNet (B0B5) as a backbone network to reduce training time and complexity. We also trained the most famous Deep Convolution neural network architectures, ResNet-152 and InceptionV3, and then compared the different architectures' accuracy results for the selected samples of the Marvel dataset. We achieved a significant improvement in classification accuracy compared to the previous state-of-the-art results for the Maritime Vessel (Marvel) dataset using the EfficientNet B5 architecture, with the highest accuracy of 91.60%.