When recognizing underwater images, problems, including poor image quality and complicated backdrops, are significant. The main problem of underwater images is the blurriness and invisibility of objects present in an image. This study presents a unique object identification design built on a YOLOv8 (You Only Look Once) framework upgraded to address these problems and further improve the models' accuracy. The study also helps in identifying underwater trash. The model is a two-phase detector model. The first phase has an Underwater Image Enhancer (UIE) data augmentation technique that works with Laplacian pyramids and gamma correctness methods to enhance the underwater images. The second phase, the proposed refined, innovative YOLOv8 model for classification purposes, takes the output from the first stage as its input. The YOLOv8 model's existing feature extractor is replaced in this study with a new feature extractor technique, HEFA, that yields superior results and better detection accuracy. The introduction of the UIE and HEFA feature extractor method represents the significant novelty of this paper. The proposed model is pruned simultaneously to eliminate unnecessary parameters and further condense the model. Pruning causes the model's accuracy to decline. Thus, the transfer learning procedure is employed to raise it. The trials’ findings show that the technique can detect objects with an accuracy of 98.5% and a mAP@50 of 98.1% and that its real-time detection speed on the GPU is double that of the YOLOv8m model's baseline performance.

在识别水下图像时，图像质量差、背景复杂等问题非常突出。水下图像的主要问题是图像中存在的物体的模糊性和不可见性。本研究提出了一种基于 YOLOv8（You Only Look Once）框架的独特对象识别设计，该框架经过升级以解决这些问题并进一步提高模型的准确性。该研究还有助于识别水下垃圾。该模型是两相检测器模型。第一阶段采用水下图像增强器（UIE）数据增强技术，与拉普拉斯金字塔和伽玛校正方法配合使用，以增强水下图像。第二阶段，提出了用于分类目的的改进的、创新的 YOLOv8 模型，将第一阶段的输出作为输入。在本研究中，YOLOv8 模型的现有特征提取器被新的特征提取器技术 HEFA 取代，该技术可产生更出色的结果和更好的检测精度。UIE和HEFA特征提取器方法的引入代表了本文的显着新颖性。同时对所提出的模型进行剪枝，以消除不必要的参数并进一步压缩模型。剪枝会导致模型的准确性下降。因此，采用迁移学习过程来提高它。试验结果表明，该技术能够以 98.5% 的准确率和 98.1% 的 mAP@50 检测物体，并且其在 GPU 上的实时检测速度是 YOLOv8m 模型基准性能的两倍。