SCR-Net: A novel lightweight aquatic biological detection network

Area of Science:

• Marine biology
• Computer vision
• Environmental monitoring

Background:

• Accurate and rapid marine biological detection is essential for conservation and resource management.
• Current methods face challenges in speed, accuracy, and resource efficiency for aquatic species identification.
• Developing lightweight yet effective detection networks is crucial for practical underwater applications.

Purpose of the Study:

• To propose a novel, fast, and efficient lightweight target detection network (SCR-Net) for marine biological detection.
• To enhance feature aggregation and spatial pooling capabilities for improved model performance.
• To reduce computational costs and parameters in feature fusion and accelerate model inference.

Main Methods:

• Implementation of a Spatial Pyramid Pool ELAN (SPPE) module combining ELAN's feature aggregation with SPPF's pooling.
• Introduction of a Cross-Scale Feature Fusion Pyramid (CFFP) structure to minimize parameters and computational load during fusion.
• Design of a lightweight feature extraction module (RGE) using low-cost processes and reparameterization for faster inference.

Main Results:

• SCR-Net demonstrates a 57.4% reduction in parameters and 37% less computation compared to baseline models.
• Achieved a mean Average Precision (mAP@0.5) of 83.2% on the DUO dataset.
• Ablation and comparative experiments confirm the effectiveness of proposed modules and superior performance over existing lightweight models on DUO and UDD datasets.

Conclusions:

• SCR-Net offers a significant advancement in lightweight underwater target detection.
• The proposed SPPE, CFFP, and RGE modules effectively enhance detection speed, efficiency, and accuracy.
• SCR-Net provides a promising solution for resource-constrained marine biological detection applications.