A Comparison of Deep Learning Approach for Underwater Object Detection

Abstract

In recent year, marine ecosystems and fisheries becomes potential resources, therefore, monitoring of these objects will be important to ensure their existence. One of computer vision techniques, it is object detection, utilized to recognize and localize objects in underwater scenery. Many studies have been conducted to investigate various deep learning methods implemented in underwater object detection; however, only a few investigations have been performed to compare mainstream object detection algorithms in these circumstances. This article examines various state-of-the-art deep learning methods applied to underwater object detection, including Faster-RCNN, SSD, RetinaNet, YOLOv3, and YOLOv4. We trained five models on RUIE dataset, then the average detection time used to compare how fast a model can detect object within an image; and mAP also applied to measured detection accuracy. All trained models have costs and benefits; SSD was fast but had poor performance; RetinaNet had consistent performance across different thresholds but the detection speed was slow; YOLOv3 was the fastest and had sufficient performance comparable with RetinaNet; YOLOv4 was good at first but performance dropped as threshold enlargement; also, YOLOv4 needed extra time to detect objects compared to YOLOv3. There are no models that are fully suited for underwater object detection; nonetheless, when the mAP and average detection time of the five models were compared, we determined that YOLOv3 is the best acceptable model among the evaluated underwater object detection models.