Incorporating Transfer Learning Strategy for improving Semantic Segmentation of Epizootic Ulcerative Syndrome Disease Using Deep Learning Model

Abstract

Automated fish disease detection can eliminate the need for manual labor and provides earlier detection of fish disease such as EUS (Epizootic Ulcerative Syndrome) before it further spreads throughout the water. One of the problems that is faced on implementing a semantic segmentation fish disease detection system is the limited size of the semantic segmentation dataset. On the other hand, classification datasets for fish disease detections are more common and available in larger sizes, which cannot be used in segmentation tasks directly since it lacks the necessary label for such tasks. In this paper, we propose a training strategy based on transfer learning to learn from both ImageNet and classification dataset before being trained on the segmentation dataset. Specifically, we first train the ImageNet pre-trained VGG16 on a classification task with the classification dataset, then we transfer the weights into a semantic segmentation architectures such as U-Net and SegNet, and finally train the segmentation network on a segmentation task with the segmentation dataset. We modify the U-Net architecture so that the pre-trained VGG16 weights can be transferred into the architecture. We used a classification dataset containing 304 images of fish diseases for classification task and a segmentation dataset containing 25 images of EUS-affected fishes for the segmentation task. The proposed training strategy is then compared with alternative training strategies such as training VGG16 on ImageNet alone or classification dataset alone. When applied to SegNet and U-Net, the proposed training strategy surpasses their respective architecture trained on ImageNet or classification dataset alone. Between these two architectures with all compared training strategies, the SegNet architecture trained with our proposed training strategy achieves the best performance with validation and testing mIoU of 66.53% and 63.46%, respectively.