Robust Underwater Image Classification Using Image Segmentation, CNN, and Dynamic ROI Approximation

Abstract

Finding classified rectangular regions of interest (ROIs) in underwater images is still a challenge, and more so if the images pose low quality with respect to illumination conditions, sharpness, and noise. These ROIs can help humans find relevant regions in the image quickly or they can be used as input for automated structural health monitoring (SHM). This task itself should be conducted automatically, e.g., used for underwater inspection. Underwater inspections of technical structures, e.g., piles of a sea mill energy harvester, typically aim to find material changes in the construction, e.g., rust or pockmark coverage, to make decisions about repair and to assess the operational safety. We propose and evaluate a hybrid approach with segmented classification using small-scaled CNN classifiers (with fewer than 20,000 hyperparameters and 3M unity vector operations) and a reconstruction of labelled ROIs by using an iterative mean and expandable bounding box algorithm. The iterative bounding box algorithm combined with bounding box overlap checking suppressed wrong spurious segment classifications and represented the best and most accurate matching ROI for a specific classification label, e.g., surfaces with pockmark coverage. The overall classification accuracy (true-positive classification) with respect to a single segment is about 70%, but with respect to the iteratively expanded ROI bounding boxes, it is about 90%.