A small underwater object detection model with enhanced feature extraction and fusion

Area of Science:

• Marine Biology
• Computer Vision
• Deep Learning

Background:

• Small object detection is vital for underwater environmental protection and marine life monitoring.
• Deep learning offers efficient detection techniques, but underwater challenges persist.
• Underwater environments pose difficulties due to complexity, limited small object data, and computational constraints.

Purpose of the Study:

• To develop an efficient deep convolutional network model for small object detection in underwater environments.
• To address the challenges of feature retention, multi-scale extraction, and computational efficiency.

Main Methods:

• Introduced a CSP for small object and lightweight (CSPSL) module to enhance feature retention.
• Proposed a variable kernel convolution (VKConv) for dynamic kernel size adjustment and multi-scale feature extraction.
• Presented a spatial pyramid pooling for multi-scale (SPPFMS) method to preserve small object features.

Main Results:

• Ablation experiments on the UDD dataset confirmed the effectiveness of the proposed methods.
• Comparative experiments on UDD and DUO datasets showed superior performance over state-of-the-art methods.
• The model achieved the best performance in computational cost and detection accuracy for real-time tasks.

Conclusions:

• The proposed deep convolutional network model effectively addresses challenges in underwater small object detection.
• The integration of CSPSL, VKConv, and SPPFMS modules significantly improves detection performance.
• The model offers a promising solution for real-time, accurate underwater small object detection.