Underwater small target detection under YOLOv8-LA model.

Area of Science:

• Marine environmental engineering
• Computer vision
• Deep learning

Background:

• Accurate underwater target detection is crucial for marine environmental engineering.
• Existing Convolutional Neural Network (CNN) models face challenges with slow processing speeds and detecting small, dense targets.
• Deep neural networks often require extensive parameters, hindering real-time applications.

Purpose of the Study:

• To propose an improved neural network model, YOLOv8-LA, for enhanced underwater target detection.
• To address limitations in processing speed and accuracy for small, densely arranged underwater targets.
• To optimize feature extraction and information aggregation for superior detection performance.

Main Methods:

• Development of a Lightweight Efficient Partial Convolution (LEPC) module for optimized spatial feature extraction and reduced computation.
• Integration of the AP-FasterNet architecture, utilizing depth-separable convolutions, to improve the capture of small target features.
• Incorporation of the content-aware reorganization (CARAFE) up-sampling operation into YOLOv8 to enhance contextual information aggregation and minimize information loss.

Main Results:

• The YOLOv8-LA model achieved 84.7% mean accuracy (mAP) on the URPC2021 dataset.
• The model demonstrated a processing speed of 189.3 frames per second (FPS) on an Nvidia GeForce RTX 3090.
• YOLOv8-LA outperformed existing state-of-the-art methods in both detection accuracy and processing speed.

Conclusions:

• The YOLOv8-LA model offers a significant advancement in underwater target detection.
• The proposed model effectively balances high detection accuracy with real-time processing capabilities.
• This research provides a robust solution for marine environmental engineering applications requiring efficient and accurate underwater surveillance.