Scholarlink

<p>Side-scan sonar systems have gained prominence in recent years, particularly following the disappearance of Malaysian Airlines Flight MH370, which necessitated extensive underwater searches to locate the aircraft’s wreckage. These efforts face considerable challenges, including the difficulty of differentiating anthropogenic structures from natural formations, limited visibility due to sediment interference, and the potential for human error during prolonged monitoring tasks. This study addresses these challenges by developing a comprehensive methodology that incorporates advanced segmentation techniques for optimal object-background separation and smear enhancement to improve image clarity. The processed images yield handcrafted features that are extracted and utilized to inform a deep convolutional neural network (CNN) designed to capture salient characteristics relevant to underwater environments. State-of-the-art methods fail to effectively capture the intricate features of underwater objects due to their reliance on predefined feature representations. By integrating our handcrafted features with a pre-trained VGG19 model, we achieved an overall validation accuracy of 99.10% after training on 70% of the dataset and evaluating it on the remaining real images, significantly surpassing the performance of previous methods. This research demonstrates that the combination of deep transfer learning and handcrafted features effectively mitigates the limitations of traditional approaches, enhancing underwater object classification and addressing the ambiguities commonly associated with conventional deep learning methodologies.</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

Multi-class classification in side-scan sonar images using handcrafted features fused with deep transfer learning

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Manishkumar Madankumar Das,
  • Soumendu Chakraborty,
  • Sooraj K. Ambat

摘要

Side-scan sonar systems have gained prominence in recent years, particularly following the disappearance of Malaysian Airlines Flight MH370, which necessitated extensive underwater searches to locate the aircraft’s wreckage. These efforts face considerable challenges, including the difficulty of differentiating anthropogenic structures from natural formations, limited visibility due to sediment interference, and the potential for human error during prolonged monitoring tasks. This study addresses these challenges by developing a comprehensive methodology that incorporates advanced segmentation techniques for optimal object-background separation and smear enhancement to improve image clarity. The processed images yield handcrafted features that are extracted and utilized to inform a deep convolutional neural network (CNN) designed to capture salient characteristics relevant to underwater environments. State-of-the-art methods fail to effectively capture the intricate features of underwater objects due to their reliance on predefined feature representations. By integrating our handcrafted features with a pre-trained VGG19 model, we achieved an overall validation accuracy of 99.10% after training on 70% of the dataset and evaluating it on the remaining real images, significantly surpassing the performance of previous methods. This research demonstrates that the combination of deep transfer learning and handcrafted features effectively mitigates the limitations of traditional approaches, enhancing underwater object classification and addressing the ambiguities commonly associated with conventional deep learning methodologies.