Deep Residual Vision Transformer Framework for Robust Underwater Object Recognition

View at publisher Summarise with VIRAT
Authors: Monica D, V Parthipan
Cited by: 0
References: 15

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Object recognition underwater is still a major issue as the visual impairment is so enormous because of light absorption, turbidity and color distortion. In order to overcome these shortcomings, this paper presents a Deep Residual Vision Transformer (DR-ViT) scheme that integrates underwater image enhancement with multi-scale residual attention to discriminate features. The dataset upon which the model was trained and tested consisted of an underwater image dataset of 3,850 annotated samples that were gathered on diverse visibility conditions. The experimental findings indicate that the proposed DR-ViT has 95.4 per cent overall accuracy which is better than CNN (82.3), ResNet50 (88.5) and baseline ViT (90.1) architectures. Moreover, the model has 96.8 and 88.1 percent accuracy in clear and high turbidity water, respectively, so it is highly robust to the change in the environment. The studies of ablation verify that residual feature refinement brings a$\mathbf{5. 8 \%}$recognition performance improvement. The suggested framework improves the underwater visual perception and enables environmentally friendly marine monitoring and autonomous operations. The paper provides a stable solution to the underwater exploration systems that need reliable recognitions in the complex water conditions.

Fields of Study & Subjects

Advanced Neural Network Applications Underwater Vehicles and Communication Systems Image Enhancement Techniques

Publication details

Journal / source
International Conference on Communication, Computing and Electronics Systems (ICCCES)
Pages
1652–1657
Publication date
21 January 2026
Language
EN