Sea Debris Detection Using Deep Learning : Diving Deep into the Sea

Abstract

Debris produced by humans is typically released into natural underwater environments like rivers and oceans. It's essential to spot marine debris in rivers and oceans so that its impact on the ecosystem can be recognised and reduced. Manually determining the amount of debris present in the oceans is time-consuming, labour-intensive, and has a limited coverage area. This paper aims at a system for the automatic detection and extraction of the debris region in an input image. We have used a deep-learning-based algorithm to perform the task of visually detecting waste in natural underwater ecosystem, with the goal of eventually deploying Autonomous Underwater Vehicles (AUVs) to explore, track, and remove that debris. The deep learning model (InceptionResNetV2 architecture) is used for training. A vast and publicly accessible dataset of real debris in assorted areas is annotated by taking it down from the JAMSTEC e-library of deep-sea debris database. The qualified network is then tested on a series of images from other parts of the dataset, yielding information on how to improve an AUV's detection capability for underwater trash removal. The trained system detects features of interest in a test image using the bounding box and class label maps. The proposed deep learning-based model architecture predicts the class labels with an accuracy of 96% and the bounding box with an accuracy of 82%. Our findings are expected to advance the goals of using AUVs to automatically survey, identify, and capture aquatic debris in underwater environments.