Scholarlink

<p>Underwater object detection is vital for marine ecosystem preservation and marine life monitoring. However, challenges arise due to the deterioration of images caused by varying illumination conditions beneath the sea. To address this, we propose a method that combines a fusion of Contrast Limited Adaptive Brightness Preserved Histogram Equalization (CLABPHE) and image sharpening for efficient image restoration. This is followed by a modified YOLOv5 model, incorporating a transformer block module and cross-convolution operation in the backbone network and neck layer, which enhances feature extraction of objects, enabling better target detection of varying-sized objects under different lighting conditions and in complex underwater environments. Here, we show that our proposed system outperforms various YOLO versions, including YOLOv5, YOLOv8, YOLOv11, YOLO-Former, TPH-YOLOv5, and Faster RCNN, with a precision increase of 2.5%, an F1 score increase of 0.9%, and an mAP value hike of 1.5% compared to the original YOLOv5. This demonstrates an effective detection of our enhancement technique, followed by the deep learning method for marine applications.</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

Transformer-enhanced YOLOv5 With cross-convolution for underwater object detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月22日

查看全文
  • Hena Prince,
  • Binesh T

摘要

Underwater object detection is vital for marine ecosystem preservation and marine life monitoring. However, challenges arise due to the deterioration of images caused by varying illumination conditions beneath the sea. To address this, we propose a method that combines a fusion of Contrast Limited Adaptive Brightness Preserved Histogram Equalization (CLABPHE) and image sharpening for efficient image restoration. This is followed by a modified YOLOv5 model, incorporating a transformer block module and cross-convolution operation in the backbone network and neck layer, which enhances feature extraction of objects, enabling better target detection of varying-sized objects under different lighting conditions and in complex underwater environments. Here, we show that our proposed system outperforms various YOLO versions, including YOLOv5, YOLOv8, YOLOv11, YOLO-Former, TPH-YOLOv5, and Faster RCNN, with a precision increase of 2.5%, an F1 score increase of 0.9%, and an mAP value hike of 1.5% compared to the original YOLOv5. This demonstrates an effective detection of our enhancement technique, followed by the deep learning method for marine applications.