INCREASING DOMAIN ADAPTABILITY FOR MACHINE LEARNING IN UNDERWATER IMAGE PROCESSING

The deep sea is one of the least explored environments on the planet. However, with technology advancing, scientists are finally able to explore these untouched areas with instrumentation such as Remote Autonomous Vehicles (ROVs) and Autonomous Underwater Vehicles (AUVs). Large volumes of data are now being collected in short amounts of time, leading to a backlog of data in need of processing. Employing the human eye to analyze such massive quantities  of video data is no longer feasible; we need an efficient alternative. Computer vision, a subfield of machine learning, is one solution to processing visual data. However, marine underwater imagery presents many challenges to machine learning, including class imbalances in datasets, variation in illumination and image quality, and an overall lack of sufficient training data. One resource that is quickly gaining momentum is FathomNet, an open source, underwater image database meant for enabling machine learning in the ocean. However, even the pre-trained models available on FathomNet experience difficulty in detecting marine objects from different domains. This study investigates novel methods for increasing the domain adaptability of the megalodon FathomNet object detector through fine-tuning processes and image enhancement algorithms. The fine-tuned megalodon model showed increased performance compared to the original megalodon model and revealed specific challenges in detecting objects in blurry underwater images. This suggests that underwater image enhancement algorithms may be employed to reduce distortion in images and improve model performance within different domains. 

Presentation Preference: Either

Primary Presenter: Kira Kaplan, Smith College (kakaplan@smith.edu)

Authors:

Kira Kaplan, Smith College  (kakaplan@smith.edu)