UAV-Based Barnacle (Cirripedia) Detection System for Dry-Docked Ship Hulls Using Transfer Learning

This study used image processing to develop a UAVbased barnacle detection system for dry-docked ship hulls. The objectives were to design and implement image processing algorithms for automated barnacle detection, analyze the system's accuracy during training and testing, conduct field tests to validate real-time effectiveness, and compare the system's performance with traditional manual inspections regarding speed and cost-effectiveness. Two models (MobileNetV1 and MobileNetV2) from Edge Impulse were trained and compared based on accuracy. The selected model, MobileNetV1, was trained and tested with an 80%/20% split of 900 datasets categorized into three classifications: areas with barnacles, clear areas, and corroded areas. Considering the unbalanced data, the Welch T-test was used to compare the performances. On-site testing validated the system's real-time effectiveness by evaluating the consistency of detection results. The UAV-based system's performance was compared with manual inspection methods using ISO/IEC 25000, highlighting its efficiency and cost-effectiveness. MobileNetV1 achieved a 91.0% accuracy during training and 90.6% during testing, which was statistically significant. The detection system demonstrated consistent confidence scores throughout on-site testing and proved to be time and cost-efficient based on ISO/IEC 25000 standards. The study suggests that machine learning could replace manual ship hull inspections due to its convenience over traditional methods.