Exploring the necessity of mosaicking for underwater imagery semantic segmentation using deep learning.

Deep learning applications are attracting considerable interest nowadays and image analysis pipelines are no exception. Benthic studies often rely on the subjective evaluation of video material recorded using underwater drones. The demand for automatic image segmentation and quantitative evaluation arises due to the large volume of video data collected. This study performed a semantic segmentation task by training the PSPNet architecture with ResNet-34 backbone for 50 epochs using imagery prepared by simply extracting a few video frames or stitch- ing a multitude of frames into a large 2D mosaic. Mosaicking is a particularly resource-intensive step, therefore, the possibility to skip such preprocessing would result in a more rapid analysis. The effect on the resulting seg- mentation quality was investigated by estimating the seabed coverage of three classes (Furcellaria lumbricalis, Mytilus edulis trossulus, and boulders) in a video material obtained from the Baltic Sea. Segmentation success, measured by intersection over union, varied between 0.56 and 0.84, usually slightly better for frames than for the mosaic overall. Absolute differences in estimated coverage were negligible (mosaic vs. frames): 0.24% vs. 1.26% for furcellaria, 0.44% vs. 2.46% for mytilus, and 4.02% vs. 2.06% for boulders. Due to the differences between predicted coverage and the mosaic-based ground truth being in an acceptable range, the findings suggest that the mosaicking step could be safely skipped in favor of a few equally spaced sample frames.