Computer Science
T. Rama M; A. V. Turukmane
Volume 23, Issue 6 , November and December 2026, , Pages 157-188
Abstract
Detecting underwater objects is crucial for scientific inquiry, resource development, and marine conservation. Current underwater object detection methods face two main challenges: (1) high parameter counts leading to excessive memory usage, and (2) reduced detection accuracy. To address ...
Read More
Detecting underwater objects is crucial for scientific inquiry, resource development, and marine conservation. Current underwater object detection methods face two main challenges: (1) high parameter counts leading to excessive memory usage, and (2) reduced detection accuracy. To address these limitations, this study introduces a computationally efficient underwater object recognition framework that integrates advanced feature extraction with fuzzy-driven adaptive classification. A novel dual-stream extractor combining Improved YOLOv9 for refined spatial localization and a Modified TResNet for enhanced semantic feature extraction is developed to capture complementary deep features. These streams are fused using Rényi entropy, ensuring that only the most informative and non-redundant representations are preserved. The fused features are then classified using a Neuro-Fuzzy Random Vector Functional Link (NF-RVFL) network, enabling fast, robust, and interpretable decision-making with minimal parameter overhead. Grad-CAM++ is incorporated to highlight critical attention regions, further improving framework transparency. Extensive experiments conducted on four public underwater datasets URPC, DLMU 2024, DUO, and RUOD show that the proposed framework consistently surpasses state-of-the-art detectors, achieving notable improvements in mAP across all benchmarks. Comprehensive ablation analysis and statistical evaluations confirm the contribution of each proposed component. Furthermore, deployment evaluation demonstrates that the proposed framework’s reduced parameter count and high efficiency make it well suited for real-time underwater platforms with limited computational resources.