An Improved Tuna-YOLO Model Based on YOLO v3 for Real-Time Tuna Detection Considering Lightweight Deployment

Yuqing Liu, Huiyong Chu, Liming Song, Zhonglin Zhang, Xing Wei, Ming Chen, Jieran ShenJournal of Marine Science and Engineering 2023 · DOI

TL;DR

This paper introduces Tuna-YOLO, a lightweight object detection model based on YOLOv3, designed for real-time tuna detection on mobile devices. By replacing the backbone with MobileNetv3, incorporating CBAM attention, and employing knowledge distillation, Tuna-YOLO significantly reduces model size and increases detection speed while maintaining high accuracy, outperforming YOLOv5 variants.

What Problem Does This Paper Address?

Accurate tuna catch statistics are crucial for fisheries management, but traditional methods are labor-intensive and costly. While deep learning offers a solution, existing object detection models are often too computationally demanding for deployment on resource-constrained mobile devices used on fishing vessels. This limits the ability to perform real-time analysis and data collection directly at sea, hindering timely stock assessment and management. The paper aims to bridge this gap by developing a highly efficient and accurate tuna detection system suitable for lightweight deployment.

Methods

The researchers built upon the YOLOv3 architecture, a well-established object detection framework, to create Tuna-YOLO. Key modifications include:

  • Lightweight Backbone: The original Darknet-53 backbone was replaced with MobileNetv3. This choice was driven by a comparative analysis of various lightweight backbones, aiming to drastically reduce the number of parameters and computational load.
  • Attention Mechanism: The SENET module was swapped for a CBAM (Convolutional Block Attention Module). This was done to enhance the model's ability to focus on relevant local features, which is critical for distinguishing tuna amidst complex backgrounds and similar-looking species.
  • Knowledge Distillation: To counteract the potential accuracy drop from using a lightweight backbone, knowledge distillation (KD) was employed. A larger, more powerful "teacher" network (DenseNet201-YOLOv3) was used to guide the training of the smaller "student" Tuna-YOLO, transferring its learned "dark knowledge" to improve the student's detection accuracy.
  • Anchor Box Optimization: A K-means clustering algorithm was used to determine optimal anchor box sizes based on the custom dataset, improving detection precision.
  • Image Augmentation: To address the low-resolution and poor lighting conditions often encountered on fishing vessels, several image enhancement techniques (saturation adjustment, histogram equalization, gamma correction, and improved multi-scale Retinex) were applied to improve image quality and aid detection.

Datasets

The study utilized a custom dataset created from electronic surveillance videos of fishing boats. This dataset was deframed and annotated, focusing on four specific tuna species: Xiphias gladius, Thunnus obesus, Thunnus albacares, and Makaira mazara. The dataset was split into training, validation, and testing sets with an 8:1:1 ratio.

Key Insights and Ideas

  • Lightweighting vs. Accuracy Trade-off is Solvable: The paper demonstrates that it's possible to significantly reduce model size and increase speed without a substantial loss in accuracy, a common challenge in deploying deep learning models on edge devices.
  • CBAM's Advantage for Tuna: The choice of CBAM over SENET highlights the importance of localized feature attention for fine-grained object detection tasks like distinguishing between similar fish species.
  • Knowledge Distillation as a Powerful Tool for Efficiency: KD proved instrumental in recovering accuracy lost during model compression, showing its effectiveness in transferring knowledge from large models to smaller, more deployable ones.
  • Custom Dataset and Augmentation are Crucial: The specific challenges of the fishing environment (low light, low resolution) necessitated a custom dataset and targeted image augmentation strategies to achieve good performance.

Key Results

The Tuna-YOLO model achieved impressive results:

  • Model Size Reduction: Parameters were reduced from 234.74 MB (YOLOv3) to 88.45 MB.
  • Speed Improvement: Detection speed increased from 10.12 fps (YOLOv3) to 15.23 fps.
  • Accuracy Enhancement: Detection precision ([email protected]) increased from 93.33% to 95.83% after knowledge distillation.
  • Performance Comparison: Tuna-YOLO outperformed YOLOv5-6.1 (small, medium, and large variants) in terms of a balance between speed and accuracy.

Limitations and Future Work

  • Dataset Size: While a custom dataset was created, its size is not explicitly detailed, and larger, more diverse datasets could further improve generalization.
  • Species Specificity: The model is trained on four specific tuna species. Future work could expand this to include a wider variety of fish species or other marine life.
  • Real-World Deployment Challenges: While designed for lightweight deployment, the paper doesn't detail the practicalities of integrating and running the model on actual onboard hardware, which might involve further optimization or hardware-specific tuning.
  • Further Optimization: The authors suggest that Tuna-YOLO provides a basis for future work, implying potential for further fine-tuning and exploration of other lightweight architectures or optimization techniques.

Summary updated

Abstract

A real-time tuna detection network on mobile devices is a common tool for accurate tuna catch statistics. However, most object detection models have multiple parameters, and normal mobile devices have difficulties in satisfying real-time detection. Based on YOLOv3, this paper proposes a Tuna-YOLO, which is a lightweight object detection network for mobile devices. Firstly, following a comparison of the performance of various lightweight backbone networks, the MobileNet v3 was used as a backbone structure to reduce the number of parameters and calculations. Secondly, the SENET module was replaced with a CBAM attention module to further improve the feature extraction ability of tuna. Then, the knowledge distillation was used to make the Tuna-YOLO detect more accurate. We created a small dataset by deframing electronic surveillance video of fishing boats and labeled the data. After data annotation on the dataset, the K-means algorithm was used to get nine better anchor boxes on the basis of label information, which was used to improve the detection precision. In addition, we compared the detection performance of the Tuna-YOLO and three versions of YOLO v5-6.1 s/m/l after image enhancement. The results show that the Tuna-YOLO reduces the parameters of YOLOv3 from 234.74 MB to 88.45 MB, increases detection precision from 93.33% to 95.83%, and increases the calculation speed from 10.12 fps to 15.23 fps. The performance of the Tuna-YOLO is better than three versions of YOLO v5-6.1 s/m/l. Tuna-YOLO provides a basis for subsequent deployment of algorithms to mobile devices and real-time catch statistics.
BibTeX
@article{liu2023improved,
  author = {Yuqing Liu and Huiyong Chu and Liming Song and Zhonglin Zhang and Xing Wei and Ming Chen and Jieran Shen},
  title = {{An Improved Tuna-YOLO Model Based on YOLO v3 for Real-Time Tuna Detection Considering Lightweight Deployment}},
  journal = {Journal of Marine Science and Engineering},
  year = {2023},
  doi = {10.3390/jmse11030542},
}
Figure