Multiclass Image Segmentation using Deep Residual Encoder-Decoder Models in Highly Turbid Underwater Ambiances

Abstract

Underwater environments, especially the coral reefs, are the habitat of many critically endangered species. Extensive monitoring of these aquatic ecosystems is essential for conserving and deep understanding of these vulnerable habitats. Monitoring by extracting details from underwater images of turbid, hazy marine environments is extremely challenging. In this work, a novel annotated dataset is created for three classes of objects in the images of coral reef environment considering fish, rock/coral and background for the Fish4Knowledge dataset, a benchmark dataset primarily for binary segmentation. This work also proposes a multiclass ResUnet based image segmentation model for the newly created multiclass annotations. Various encoder-decoder convolutional architectures were analysed and found that ResUnet exhibits better robustness. The performance of the multiclass ResUnet model is also analysed by optimizing with different cost functions. Various underwater noisy conditions are simulated in the test images to find the robustness of the model, and observed that the proposed model optimised with Jaccard loss performs better even in extremely noisy scenarios.

Keywords

• underwater image
• encoder-decoder model
• multi-class image segmentation
• fish4knowledge
• underwater noise