TL;DR
This paper introduces a novel two-step image processing technique to enhance underwater crab images, addressing issues like low brightness, fog, and color imbalance. By combining improved dark channel prior defogging with a unique color correction method, they significantly improve image quality. This enhanced imagery is then fed into an optimized YOLOv5s model, achieving a remarkable 90.78% detection rate for crabs, paving the way for automated aquaculture management.
What Problem Does This Paper Address?
Crab aquaculture is a major industry in China, but current methods for monitoring crab populations rely on inefficient manual counting. Underwater images, crucial for automated monitoring, are often plagued by poor visibility due to depth, low light, and water turbidity, resulting in low contrast, fog, and color distortion. These issues make it difficult for even advanced machine vision systems to accurately detect and count crabs. This paper aims to overcome these challenges by developing a robust underwater image processing pipeline that enhances image quality, enabling more accurate and efficient automated crab detection.
Methods
The proposed solution involves a two-stage approach:
- Underwater Image Enhancement:
- Improved Dark Channel Prior (DCP): The standard DCP algorithm is enhanced by optimizing its parameters (transmittance and atmospheric light) using guided filtering and quadtrees. This helps to remove fog while preserving image details and mitigating issues caused by artificial lighting.
- Two-Step Color Correction: * First Step: Combines the "gray world assumption" and "perfect reflection assumption" to correct the red channel, aiming to balance colors.
- Second Step: Employs a "strong channel compensation for weak channel" strategy, specifically using the green channel to enhance the red and blue channels. This further refines color balance and enriches image information.
- Optimized Target Detection:
- Lightweight Backbone: The original YOLOv5s backbone is replaced with ShuffleNetV2. This significantly reduces model size and computational cost, leading to faster inference speeds without a substantial loss in accuracy.
- Anchor Box Optimization: A K-means++ clustering algorithm is used to generate custom anchor boxes tailored to the underwater crab dataset, improving the initial bounding box predictions.
- Loss Function: DIoU loss is employed to enhance bounding box regression, leading to more accurate localization.
Datasets
The study utilized a custom dataset of 3000 underwater crab images collected from a crab farming demonstration base in Changzhou City, Jiangsu Province, China. These images were captured using underwater cameras with auxiliary low-light lighting. The dataset was divided into training and testing sets in a 7:3 ratio.
Key Insights and Ideas
- DCP isn't enough for underwater: The paper highlights that standard DCP, while effective for general fog removal, struggles with the unique challenges of underwater environments, particularly artificial lighting and color casts.
- Green channel as a savior: The innovative use of the green channel to compensate for deficiencies in the red and blue channels is a key insight for correcting underwater color imbalance.
- Lightweight models are crucial for real-time applications: Replacing the YOLOv5s backbone with ShuffleNetV2 demonstrates that achieving high detection rates doesn't necessarily require massive models, making real-time deployment feasible.
- Custom anchor boxes matter: Tailoring anchor boxes to the specific dataset (underwater crabs) significantly improves detection performance compared to generic datasets like COCO.
Key Results
- Detection Rate: The proposed method achieved a detection rate of 90.78%. This is a significant improvement over the original image detection rate of 74.72%.
- Average Confidence Level: The average confidence level was 0.75, a substantial increase from 0.51 in the original images.
- Performance Gains: Compared to the original YOLOv5s detection, the proposed method showed a 21.41% increase in detection rate and a 47.06% increase in average confidence level.
- Objective Metrics: The proposed method consistently ranked first in objective evaluations like SSIM (mean value of 0.864) and PSNR, indicating superior image quality preservation and enhancement.
- Speed: The optimized YOLOv5s model with ShuffleNetV2 achieved 50 FPS, a significant improvement for real-time applications.
Limitations and Future Work
The authors acknowledge a few limitations:
- Dataset Specificity: The algorithm was developed and tested on a specific crab dataset. Its generalization to other underwater environments or species might be limited.
- Artificial Light Artifacts: While addressed, the influence of artificial light sources can still lead to "too-vivid color blocks" in some processed images, affecting visual experience.
Future work suggested includes:
- Expanding the Dataset: To improve robustness and handle variations in crab morphology and shooting angles, a larger and more diverse dataset is needed.
- Building a Real-time System: The ultimate goal is to deploy this image processing and detection algorithm on an embedded board for an automated baiting boat, enabling real-time monitoring and intelligent aquaculture management.
- Integrating Environmental Data: Combining crab distribution maps with other environmental data (water quality, pond area) for more precise baiting strategies.
Summary updated