光学图像目标检测技术是水下目标检测技术的重要分支，是海洋近距离探索中最为重要和可靠的手段。随着深度学习技术的发展，基于深度学习网络进行端到端训练的检测方法在自然图像目标检测任务中得到了广泛的运用。但是在水下图像目标检测任务中，由于水对于光严重的散射和吸收效应，水下图像的质量退化严重，给水下光学图像目标检测带来了困难。现有的方法试图通过端到端训练的方式来让深度学习网络学会处理退化严重的水下图像，但是却因为缺乏水下光学成像的物理模型知识导致效果不佳。 本论文将水下成像过程的物理模型知识和深度学习网络相结合，提出了一个基于物理模型的水下光学图像目标检测方案来解决水下图像质量退化的问题。水下图像的目标检测过程被分解为3个步骤，分别为水下图像的分割与分类、目标图像块的增强和图像块中的目标定位。（1）针对水下图像的分割与分类任务，提出了基于候选区域生成和特征共享结构深度学习分类网络，实现对整张水下图像一次特征提取即可完成所有前景图像块的提取和分类。（2）针对水下目标图像块增强任务，提出了基于水下光场散射规律物理模型的水下图像去雾和迭代去模糊方法。（3）针对图像块中的目标定位任务，提出的基于特征金字塔结构和动态加权损失函数的深度学习定位网络，从而有效实现了对于分割增强后的小图像块中目标的定位。 实验结果表明，在实际水下图像数据集URPC上，论文提出的基于物理模型的水下光学图像目标检测方案能够将目标检测性能指标mAP从56.5提升至58.1，验证了方案的有效性。

Optical image object detection technology is an important branch of underwater object detection technology and the most important and reliable means in close-range ocean exploration. With the rapid development of deep learning technology, deep learning networks using end-to-end training have been widely used in natural image object detection tasks. However, in underwater image object detection scenarios, due to the severe scattering and absorption effects of water on light, the quality of underwater images is seriously degraded, which brings difficulties to underwater optical image object detection. Existing methods try to use end-to-end training to let deep learning networks learn to process severely degraded underwater images, but the results are poor due to the lack of physical model knowledge of underwater optical imaging. This thesis combines the physical model knowledge of the underwater imaging process with the deep learning network, and proposes an underwater optical image object detection solution based on the physical model to solve the problem of underwater image quality degradation. The object detection process of underwater images is decomposed into three steps, namely segmentation and classification of underwater images, enhancement of object image blocks, and in image blocks. (1) For the task of segmentation and classification of underwater images, a deep learning classification network based on candidate region generation and feature sharing structure is proposed, which can complete the extraction and classification of all foreground image blocks with one feature extraction of the entire underwater image. (2) For the task of underwater object image block enhancement, an underwater image defogging and iterative deblurring method based on the physical model of underwater light field scattering laws is proposed. (3) For the object localizing task in image blocks, a deep learning positioning network based on feature pyramid structure and dynamic weighted loss function is proposed, thus effectively achieving the localizing of targets in small image blocks after segmentation and enhancement. Experimental results show that on the actual underwater image dataset URPC, the underwater optical image object detection scheme based on the physical model proposed in the thesis can increase the object detection performance index mAP from 56.5 to 58.1, verifying the effectiveness of the scheme.