Multi-class Segmentation of Trash in Coastal Areas Using Encoder-Decoder Architecture

Machine Learning Techniques for Smart City Applications: Trends and Solutions. Advances in Science, Technology & Innovation. Springer.

Trash accumulation in beaches affects the ecosystem of coastal lines. Different types of trash can affect beaches in a variety of ways and manual identification of this trash might become laborious. So, it becomes important to devise a method to facilitate the localization of this trash without human intervention. In this paper, we propose a deep learning-based solution for multi-class segmentation of trash objects using Unmanned Aerial Vehicles. But the problem with the aggregated orthoimages is the low foreground-to-background ratio which tends to render very high false positives (FP) during classification. To counter this, we propose a random data generation method to generate synthetic data over real backgrounds. The best performing model among the chosen candidate architectures (U-Net and SegNet) on both the real and synthetic datasets are evaluated and results are compared using various segmentation metrics. Later the segmentation masks are transposed onto a map to facilitate localization.