Semantic Segmentation of Underwater Images using UNet architecture based Deep Convolutional Encoder Decoder Model

Abstract

With rapid development of deep neural network-based image processing techniques, many areas like augmented reality, robotic vision are adopting image segmentation technique for addressing image analytics problems. Efforts are also being made in exploring underwater environments which are precious with vast biological gene banks and marine biological resources. Study of these forms of ecosystem helps human in interpreting environmental effects such as climatic change, effect of pollution and so on. Deep convolutional neural networks based architectures have paved rapid development in the field of segmentation. This paper tries to explore underwater objects by applying semantic image segmentation. Proposed model adapted UNet based semantic segmentation network as the basic frame work and extended it for accurate segmentation of underwater fish4knowledge image dataset. Training and fine tuning the model by choosing appropriate hyperparameters achieved an average IoU scores of 0.8583 on Acanthurus nigrofuscus class 08 from Fish4Knowledge image dataset.