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