Underwater Object Detection Through Analysis and Data Augmentation of Underwater Datasets

Nowadays, with the rapid development of artificial intelligence (AI) technologies, more and more jobs can be fulfilled by underwater drones. Particularly, object detection with deep learning becomes an important issue which can significantly improve the performance of underwater drones. In this work, we analyzed and augmented an existing underwater dataset and proposed a new object detection dataset for inshore aquaculture. The renewal dataset is used for the training of the network for real-time object detection. The simulation results showed that the existing underwater dataset has a class imbalance problem. In addition, it is shown that although large objects are detected with high accuracy, rocks and small shadows are sometimes mis-detected as targets. It is due to the distance between the target object and the camera lenses, simulation of the background, and the object detection model not learning the difference between background and target. This indicates that object detection accuracy could be improved by the dataset that can accurately learn object features.