Underwater image classification and object detection using shuffle convolutional channel enhancement

The complexity of the underwater environment poses significant challenges to underwater visual fields such as underwater object classification and detection, due to factors such as light attenuation, colour distortion, blurriness, noise, and complex underwater backgrounds. This paper proposes a local attention algorithm using Shuffle Convolutional Channel enhancement (SF-NN), which improves the accuracy of key feature identification in images.The SF-NN technique effectively addresses variations in the scale of visual entities and high-resolution issues, improving connectivity between windows and significantly enhancing the accuracy of underwater image classification,and it was verified on the ImageNet and FishNet datasets. Furthermore, by integrating the proposed algorithm as a feature extraction method into the Mask R-CNN framework, we effectively tackle problems such as object occlusion and deformation in complex underwater scenes, thereby improving the precision and robustness of the Mask R-CNN object detection algorithm.It has been validated on the COCO and TransCan datasets, demonstrating the significant advantages of the algorithm in improving underwater object detection performance through applications in different underwater environments and targets.

水下环境的复杂性为水下视觉任务（如水下目标分类与检测）带来了严峻挑战，这是由于光线衰减、色彩失真、图像模糊、噪声干扰以及复杂水下背景等诸多因素所致。本文提出一种采用混洗卷积通道增强（Shuffle Convolutional Channel Enhancement）技术的局部注意力算法（SF-NN），可提升图像关键特征的识别准确率。SF-NN能够有效应对视觉实体的尺度变化与高分辨率难题，增强窗口间的连通性，显著提升水下图像分类的准确率，该算法已在ImageNet与FishNet数据集上完成验证。进一步地，本文将所提算法作为特征提取方法集成至Mask R-CNN框架中，可有效解决复杂水下场景下的目标遮挡与形变等问题，进而提升Mask R-CNN目标检测算法的精度与鲁棒性。该方法已在COCO与TransCan数据集上得到验证，结果表明，通过在不同水下环境与目标场景中应用本算法，其在提升水下目标检测性能方面具备显著优势。