Underwater U-Net: Deep Learning with U-Net for Visual Underwater Moving Object detection

Underwater object detection is a prevalent way of detecting the movements of objects in the video which has created a dynamic variation. The nature of the problem is complex due to many factors including camera jitter, high haze, dynamic background, absorbing, scattering, and attenuation of light rays in the water medium. It is found that U-Net architecture is popular in image segmentation but not explored for underwater object detection. In this article, a unique attempt is made the first time to use a modified U-Net architecture with ResNet encoders to detect underwater objects. The proposed object detection scheme relies on the principle of background subtraction or local change detection, where a robust background is constructed by taking multiple image frames of a sequence in deeper high dimensional space. The proposed scheme is tested on four different databases and for page constraint, the results are reported here on UNICT underwater database. The objective evaluation of the proposed scheme is carried out using four different performance evaluation measures: precision, recall, F-measures, and similarity. Both the visual and objective evaluation corroborates our findings.