Abstract: Underwater object detection and classification are essential for marine exploration, environmental monitoring, and autonomous underwater systems. However, complex underwater conditions such as poor visibility, color distortion, and frequent object overlap significantly degrade the performance of conventional detection methods. In particular, overlapping and occluded objects pose serious challenges to accurate localization and classification, necessitating more robust vision-based solutions. To address this problem, this work proposes an occlusion-resilient instance segmentation framework that integrates deep learning models for effective separation and identification of overlapping underwater objects. The solution combines spatial feature extraction and contextual learning to enhance boundary delineation and class discrimination under partial visibility conditions. The methodology employs the Underwater Object Detection Dataset, consisting of 10,000 annotated images across fish, coral, and debris categories. Image enhancement, normalization, augmentation, and refined instance annotations are applied, followed by CNN, RNN, and hybrid CNN–RNN modeling for segmentation and classification. Experimental results demonstrate that the proposed hybrid model achieves 96% accuracy, 94% precision, 93% recall, and an F1-score of 94%, outperforming standalone CNN and RNN architectures in handling overlapping and occluded underwater objects.

Keywords: Underwater object detection, instance segmentation, occlusion handling, deep learning, CNN–RNN hybrid model, marine image analysis.

Title: Occlusion-Resilient Instance Segmentation for Overlapping Underwater Object Detection and Classification

Author: Anuspa Sahani, Dr. Anupama Sahu, Dr. Sasmita Mishra

International Journal of Novel Research in Engineering and Science

ISSN 2394-7349

Vol. 13, Issue 1, March 2026 - August 2026

Page No: 1-16

Novelty Journals

Website: www.noveltyjournals.com

Published Date: 04-April-2026