Improving Buoy Detection with Deep Transfer Learning for Mussel Farm Automation

TL;DR

This work addresses automated buoy detection for mussel farm monitoring to reduce manual inspection costs. It adopts transfer learning by fine-tuning YOLOv7/YOLOv7-tiny models on a multi-source buoy dataset, including an adverse-weather test set, to achieve robust real-time detection with a single buoy class. Results show that 1280-pixel input models provide the best balance of $mAP$ and speed, with the full-1280 and tiny-1280 variants outperforming smaller inputs in detecting small buoys while maintaining real-time performance. The approach supports integration into automated mussel-farm pipelines for detecting drifting or sinking buoys, improving operational efficiency and resilience to challenging conditions.

Abstract

The aquaculture sector in New Zealand is experiencing rapid expansion, with a particular emphasis on mussel exports. As the demands of mussel farming operations continue to evolve, the integration of artificial intelligence and computer vision techniques, such as intelligent object detection, is emerging as an effective approach to enhance operational efficiency. This study delves into advancing buoy detection by leveraging deep learning methodologies for intelligent mussel farm monitoring and management. The primary objective centers on improving accuracy and robustness in detecting buoys across a spectrum of real-world scenarios. A diverse dataset sourced from mussel farms is captured and labeled for training, encompassing imagery taken from cameras mounted on both floating platforms and traversing vessels, capturing various lighting and weather conditions. To establish an effective deep learning model for buoy detection with a limited number of labeled data, we employ transfer learning techniques. This involves adapting a pre-trained object detection model to create a specialized deep learning buoy detection model. We explore different pre-trained models, including YOLO and its variants, alongside data diversity to investigate their effects on model performance. Our investigation demonstrates a significant enhancement in buoy detection performance through deep learning, accompanied by improved generalization across diverse weather conditions, highlighting the practical effectiveness of our approach.

Figures (5)

• Figure 1: Raw images are collected from the cameras, then annotated with buoy bounding boxes. Image augmentations are applied to the image dataset before fine-tuning the pre-trained model, resulting in our buoy detection model.
• Figure 2: Image samples. Row 1: Boat camera. Row 2: Low-res buoy-mounted camera. Row 3: High-res buoy-mounted camera. Row 4: Adverse weather low-res buoy-mounted camera.
• Figure 3: Each model variation shows mAP, mAP@0.5, and inference FPS. Higher is better for all displayed metrics.
• Figure 4: A different model is shown in each column: a. tiny-640, b. full-640, c. tiny-1280, d. full-1280. Various example images demonstrating a variety of scenarios are shown in each row as: 1. Many small buoys, 2. Choppy water, 3. Foggy conditions, 4. Rough conditions, 5. Calm conditions with distinct shadows on the water, 6. Calm conditions
• Figure 5: YOLOv7 with 1280px input model trained and tested on various datasets and also tested on the adverse weather dataset shows mAP@[0.5:0.05:0.95].