Publication

Abstract :

Identification of naval mines is a very tedious task due to turbulent nature in ocean and the varying deployment of mines, which differ by location and type, such as buoyant mines, moored mines, and seabed or ground mines. Recently, there has been growing interest in utilizing computer vision and deep learning algorithms for underwater detection tasks, including the identification of explosive targets and the monitoring of marine organisms using autonomous underwater vehicles. In this study, we conducted experiments in a lake environment, capturing optical camera data of a designed spherical mine-like target. We evaluated the performance of DenseNet-264 and CNN deep learning classification algorithms, utilizing both RMSProp and Adam optimizers for target classification. The CNN model with the Adam optimizer demonstrated superior classification performance, achieving 100% accuracy, a prediction time of 98ms, and a test loss of 9.2 x 10-8. This work can be further extended to detect targets with bounding boxes in the images and videos using NVDIA AI Edge Device for post-processing purposes.