Scholarlink

Sonar imaging system has been considered a promising solution to estimate fish populations because it can withstand murky waters and require low-cost acquisition. However, estimating the number of fish in sonar images is still challenging as sonar images contain noise or have poor resolution. In this study, ensemble method is proposed to improve the accuracy in estimating the number of fish in sonar images. In the proposed method, two deep learning models are combined by calculating the average of two outputs after each of them with ImageNet initialization is trained sequentially on DeepFish and Sonar datasets. Combining two deep learning models helps compensate for the weakness of each other, leading to performance improvement. Experimental results on sonar dataset demonstrate that the combination of VGG16 and RegNetY reduces mean absolute error, root mean square error by 3.9%, 14.2%, respectively, and improves coefficient of determination by 1.2% in comparison with individual models.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Counting Fish in Sonar Images Using Ensemble Learning Technique

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tri-Nhan Nguyen,
  • Trung-Quan Hoang,
  • Duc-Minh Nguyen,
  • Quoc-Trinh Vo

摘要

Sonar imaging system has been considered a promising solution to estimate fish populations because it can withstand murky waters and require low-cost acquisition. However, estimating the number of fish in sonar images is still challenging as sonar images contain noise or have poor resolution. In this study, ensemble method is proposed to improve the accuracy in estimating the number of fish in sonar images. In the proposed method, two deep learning models are combined by calculating the average of two outputs after each of them with ImageNet initialization is trained sequentially on DeepFish and Sonar datasets. Combining two deep learning models helps compensate for the weakness of each other, leading to performance improvement. Experimental results on sonar dataset demonstrate that the combination of VGG16 and RegNetY reduces mean absolute error, root mean square error by 3.9%, 14.2%, respectively, and improves coefficient of determination by 1.2% in comparison with individual models.