Research on Optimization Method of Multi-scale Fish Target Fast Detection Network

Yang Liu, Shengmao Zhang, Fei Wang, Wei Fan, Guohua Zou, Bo Jing

Abstract

Real-time fish target detection on embedded devices faces a significant challenge in balancing computational speed with identification accuracy. This research addresses this by introducing the "Aquarium Fish" dataset—a specialized collection of 10,042 images across 84 species—and proposing the BTP-yoloV3 network. By utilizing Depthwise convolutions to redesign the backbone, the authors reduced computational complexity by 94.1%. Through the application of data augmentation (Mosaic, CutMix, MixUp) and optimized activation functions (Mish, Swish, ELU), the final model achieved a test accuracy of 94.37% with a complexity of only 5.47 BFLOPS, providing a high-efficiency solution for practical underwater vision applications.

Introduction

The demand for automated fish monitoring in the aquarium industry and marine conservation is growing. However, standard object detection models are often too computationally heavy for the low-power embedded devices used in these environments. This paper presents an optimized approach that minimizes hardware requirements while maintaining the high precision necessary to distinguish between similar fish species.

The "Aquarium Fish" Dataset

A significant hurdle in specialized computer vision is the lack of high-quality, domain-specific data. The authors developed the "Aquarium Fish" dataset to fill this gap. Containing over 10,000 annotated images, it provides a robust benchmark for fine-grained fish detection tasks.

Overview of the Aquarium Fish dataset (a) and sample detection results (b).

Methodology: BTP-yoloV3 and TPDP

The core innovation is the TPDP (Three PDP) module, which leverages Depthwise Separable Convolutions. This structural change allows the network to filter and combine features with far fewer parameters than standard convolutions. The BTP-yoloV3 architecture integrates these modules into a multi-scale framework, ensuring effective detection across various target sizes.

Optimization Strategies

To maximize performance, the researchers employed a "bag of tricks" during training:

  • Data Augmentation: The Mosaic method, which combines four training images into one, was particularly effective in teaching the model to recognize targets at different scales.
  • Activation Functions: Replacing ReLU with smoother functions like Mish improved gradient flow and boosted final accuracy by 0.76%.
Visual comparison of data augmentation techniques: (a) Mosaic, (b) Cutmix, and (c) Mixup.

Results and Discussion

BTP-yoloV3 was evaluated against industry standards including YOLOv4 and YOLOv3-tiny. The results demonstrate that BTP-yoloV3 offers a superior balance of speed and accuracy for edge deployment.

Model Weight Size mAP (Accuracy) BFLOPS (Complexity) FPS (Speed)
YOLOv4 258M 95.40% 91.13 87.0
YOLOv3-tiny 34M 93.34% 5.577 250.0
BTP-yoloV3 13.9M 94.37% 5.476 285.7

While YOLOv4 remains the most accurate, BTP-yoloV3 achieves comparable results with a fraction of the computational cost and a significantly higher frame rate (285.7 FPS on a Tesla V100).

Conclusion

The BTP-yoloV3 network represents a significant step toward practical, real-time underwater monitoring. By combining efficient architectural design with advanced training augmentations, the researchers have created a model that is both lightweight enough for embedded hardware and precise enough for scientific and commercial use.

Summary updated

Abstract

The fish target detection algorithm lacks a good quality data set, and the algorithm achieves real-time detection with lower power consumption on embedded devices, and it is difficult to balance the calculation speed and identification ability. To this end, this paper collected and annotated a data set named "Aquarium Fish" of 84 fishes containing 10042 images, and based on this data set, proposed a multi-scale input fast fish target detection network (BTP-yoloV3) and its optimization method. The experiment uses Depthwise convolution to redesign the backbone of the yoloV4 network, which reduces the amount of calculation by 94.1%, and the test accuracy is 92.34%. Then, the training model is enhanced with MixUp, CutMix, and mosaic to increase the test accuracy by 1.27%; Finally, use the mish, swish, and ELU activation functions to increase the test accuracy by 0.76%. As a result, the accuracy of testing the network with 2000 fish images reached 94.37%, and the computational complexity of the network BFLOPS was only 5.47. Comparing the YoloV3~4, MobileNetV2-yoloV3, and YoloV3-tiny networks of migration learning on this data set. The results show that BTP-Yolov3 has smaller model parameters, faster calculation speed, and lower energy consumption during operation while ensuring the calculation accuracy. It provides a certain reference value for the practical application of neural network.
BibTeX
@misc{liu2021research,
  author = {Yang Liu and Shengmao Zhang and Fei Wang and Wei Fan and Guohua Zou and Bo Jing},
  title = {{Research on Optimization Method of Multi-scale Fish Target Fast Detection Network}},
  year = {2021},
  eprint = {2104.05050},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.05050},
}
Figure