Automatic Segmentation of Underwater Images with Shannon's Thresholding and UNet Variants

Automatic detection of the Region of Interest (RoI) from the chosen digital image is one of the prime tasks in various domains, and this work considered the underwater images for the study. In order to extract the RoI with better accuracy, the proposed research aims to develop a scheme by integrating the image enhancement approach and the Convolutional Neural Network (CNN) segmentation. The various phases of this scheme involve; (i) Image collection and pre-processing, (ii) RoI extraction using CNN segmentation, and (iii) Comparing the segmented region with Ground-Truth (GT) and confirming the merit of the implemented technique. This work considered Shannon's Entropy and Firefly Algorithm (SE+FA) tri-level thresholding to pre-process the images. Then the necessary RoI from these images are mined using the UNet and its variants, such as UNet+, UNet++, and VGG-UNet. The experimental investigation of this scheme confirms that the VGG-UNet implemented using VGG16 as the backbone (encoder) helped achieve a better result than other schemes. This investigation considered the UFO-120 database for the assessment, and the VGG-UNet offered an RoI mining accuracy of>98%. Other UNet variants offered >96% accuracy, which confirms the merit of the proposed scheme.