Scholarlink

<p>The world’s ocean depths hide a great mystery, and interpreting it will require difficult navigation. The exploration of the underwater environment has recently increased with the development of robotics and computer vision devices. Several underwater sensors have gathered vast amounts of data. Still, there are several issues, including poor image quality, challenges getting training samples, and uncontrollable objects in the water. The time required to process many images will be comparatively high and error-prone when these pictures are processed using machine learning techniques requiring human intervention. To address the limitations, we proposed a novel, efficient deep learning-based ensemble approach, Enhanced YOLOv8 and Adaptive Stacked Residual Convolutional Neural Network (ACLS-RCNN) for underwater detection and classification. In preprocessing, the clarity of the image is enhanced, and the noise is reduced by using Adaptively Clipped Contrast Limited Adaptive Histogram Equalization (ACCLAHE) and Swin-Conv-UNet approach. Features like shape, colour, and texture are extracted using a Novel ResNeXt50 approach. With the help of the approach, the essential features are extracted. Then, the Boosted Fuzzy-Based Hungry Games Search Algorithm (BFHGS) is employed to select the significant features. Features improve detection efficiency and accuracy by selecting and extracting pertinent information and lowering computational complexity. Additionally, we used the Self-Paced Ensemble and Auxiliary Classifier Generative Adversarial Networks (SPE-ACGAN) approach to tackle class imbalance issues. The experimental results show that the proposed framework yielded the best precision, recall, and mAP values when applied to the four data sets. Furthermore, the proposed approach outperforms the other existing forecasting methods analysed in this research in terms of precision (98.97%), recall (99.03%), and mAP (99.02%).</p>

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

Ocean scope: a boosted novel ensemble deep learning-driven accurate underwater object detection and classification system using enhanced YOLOv8 and ACLS-RCNN

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Jainabbi Banda,
  • J. Harikiran

摘要

The world’s ocean depths hide a great mystery, and interpreting it will require difficult navigation. The exploration of the underwater environment has recently increased with the development of robotics and computer vision devices. Several underwater sensors have gathered vast amounts of data. Still, there are several issues, including poor image quality, challenges getting training samples, and uncontrollable objects in the water. The time required to process many images will be comparatively high and error-prone when these pictures are processed using machine learning techniques requiring human intervention. To address the limitations, we proposed a novel, efficient deep learning-based ensemble approach, Enhanced YOLOv8 and Adaptive Stacked Residual Convolutional Neural Network (ACLS-RCNN) for underwater detection and classification. In preprocessing, the clarity of the image is enhanced, and the noise is reduced by using Adaptively Clipped Contrast Limited Adaptive Histogram Equalization (ACCLAHE) and Swin-Conv-UNet approach. Features like shape, colour, and texture are extracted using a Novel ResNeXt50 approach. With the help of the approach, the essential features are extracted. Then, the Boosted Fuzzy-Based Hungry Games Search Algorithm (BFHGS) is employed to select the significant features. Features improve detection efficiency and accuracy by selecting and extracting pertinent information and lowering computational complexity. Additionally, we used the Self-Paced Ensemble and Auxiliary Classifier Generative Adversarial Networks (SPE-ACGAN) approach to tackle class imbalance issues. The experimental results show that the proposed framework yielded the best precision, recall, and mAP values when applied to the four data sets. Furthermore, the proposed approach outperforms the other existing forecasting methods analysed in this research in terms of precision (98.97%), recall (99.03%), and mAP (99.02%).