TL;DR
This paper presents an improved Faster R-CNN model for accurately detecting fouling shellfish on marine aquaculture cages. By enhancing the backbone network with ResNeXt50 and CBAM, optimizing anchor boxes, and incorporating underwater image enhancement, the model achieves a remarkable 94.27% average precision while being significantly more lightweight than its predecessor.
What Problem Does This Paper Address?
Marine aquaculture is crucial for food security, but submerged cages accumulate biofouling organisms, hindering growth and water exchange. Manual cleaning is inefficient. Automating this process with Autonomous Underwater Vehicles (AUVs) requires robust detection of these fouling organisms. However, underwater environments present challenges like low visibility, color distortion, and scattering, making precise detection difficult for current AI models. This paper aims to overcome these challenges by developing a more accurate and efficient detection system for fouling shellfish on aquaculture cages.
Methods
The core of this research is an improved Faster R-CNN architecture. Key modifications include:
- Backbone Replacement: The original VGG16 backbone is replaced with ResNeXt50. This choice is motivated by ResNeXt50's ability to handle deeper networks without vanishing gradients and its more efficient parameter usage, making it suitable for AUV deployment.
- Attention Mechanism: A Convolutional Block Attention Module (CBAM) is integrated into the ResNeXt50 backbone. CBAM enhances feature extraction by focusing on both channel and spatial dimensions, allowing the model to better identify small and occluded targets.
- Anchor Box Optimization: The dimensions and aspect ratios of the anchor boxes within the Region Proposal Network (RPN) are adjusted. This is crucial for better matching the diverse shapes and sizes of fouling shellfish.
- IoU Threshold Tuning: The Intersection over Union (IoU) threshold for Non-Maximum Suppression (NMS) is optimized. This helps in filtering out redundant bounding boxes and improving detection accuracy.
- Image Enhancement: A Grayscale World Assumption (GWA) combined with Multi-Scale Retinex with Single Scale Component and Color Restoration (MSRCR) algorithm (GWA-MSRCR) is employed to preprocess underwater images. This addresses issues like underexposure, low contrast, and color bias, leading to clearer images for training.
Datasets
The study utilized a custom dataset of 800 RGB images captured by an underwater drone of fouling shellfish (primarily mussels and oysters) on marine aquaculture cages. This dataset was then augmented through resizing, random flipping, and rotation, resulting in a final training dataset of 4000 images. The images were annotated in the PASCAL VOC dataset format.
Key Insights and Ideas
- ResNeXt50 + CBAM is a powerful combination: Replacing VGG16 with ResNeXt50 and integrating CBAM significantly boosts feature extraction capabilities, especially for complex underwater scenes with small, occluded targets.
- Underwater image enhancement is critical: The GWA-MSRCR algorithm dramatically improves image quality, leading to a substantial increase in detection accuracy (11.72% higher AP) when used for training data. This highlights that addressing the visual challenges of the underwater environment is as important as the detection model itself.
- Anchor box and IoU tuning matters: Fine-tuning anchor box parameters and the IoU threshold for NMS is not just a minor adjustment; it directly impacts the model's ability to accurately localize targets, especially in challenging environments.
- Lightweight and accurate is achievable: The improved model achieves high accuracy while being significantly smaller (one-fifth the weight) than the original Faster R-CNN, making it more practical for deployment on resource-constrained AUVs.
Key Results
- The enhanced model achieved an average precision of 94.27%, a 10.31% improvement over the original model.
- The model's weight is approximately one-fifth of the original Faster R-CNN model.
- At an IoU of 0.5, the model achieved a mean average precision (mAP) of 93.14%, outperforming many prevalent detection models.
- Training with the image-enhanced dataset resulted in an 11.72 percentage point higher average precision compared to training with the original dataset.
- The optimal IoU threshold for detection was found to be 0.6, yielding the highest average detection rate of 94.27%.
Limitations and Future Work
- Water Turbidity: The authors acknowledge that water turbidity significantly impacts detection performance.
- Limited Scope: The current research focuses solely on detecting shellfish biofouling, with future work planned to explore multi-sensor fusion (optical and acoustic) for more comprehensive underwater detection.
- Dataset Size: While augmented, the dataset could potentially benefit from further expansion with more diverse underwater conditions.
- Future Work: The authors propose integrating multi-sensor fusion (optical and acoustic) for enhanced detection and collecting additional datasets to train underwater net detection models for broader applications. They also aim to explore the relationship between biofouling growth and temperature/climate conditions in subsequent research.
Summary updated