【作者】 王德新；

【导师】 崔学荣；

【作者基本信息】 中国石油大学(华东) ， 计算机科学与技术， 2023， 硕士

【摘要】 基于视觉的水下目标识别与追踪技术作为海洋科学研究的一个重要领域,可以实时监测海洋生物的分布和行为,对于保护海洋环境和生态具有重要意义。但是,复杂多变的水下环境为海洋生物的识别与追踪带来了困难与挑战,原始水下目标图像蓝绿偏色严重、对比度低、细节模糊,现有算法存在识别效率低、追踪目标易丢失等问题,因此本文展开了基于深度学习的水下图像增强、目标识别与追踪算法的研究,具体内容如下:（1）针对水下图像存在蓝绿偏色、对比度低和细节模糊等问题,提出一种基于UWNET网络的水下图像增强算法。首先,由于缺乏高质量的配对数据集,采用CycleGAN（Cycle-Consistent Generative Adversarial Networks）网络将空中图像渲染为水下风格的图像,将其作为水下图像增强网络的输入;其次,设计轻量化的水下图像增强网络UWNET,通过特征提取网络提取水下图像的通用特征,利用注意力机制将水下图像中的目标信息与背景信息进行差异化增强;最后,基于L₁绝对误差和结构相似度设计损失函数。实验结果表明,UWNET通过最小化水下图像和目标图像之间的损失解决了蓝绿偏色问题,提高了对比度,丰富了特征点数量,与现有算法相比综合性能更好。（2）针对现有算法在识别水下生物时检测速度难以满足实时检测这一问题,以YOLOv5s为基础网络进行改进,提出一种轻量化的目标检测算法M＿YOLOv5s。首先,采用深度可分离卷积和SE（Squeeze and Excitation）注意力机制模块设计主干网络,提取水下图像的特征;其次,使用特征融合能力更强的Bi FPN模块融合深层特征图中的语义信息和浅层特征图中的纹理信息,提升小目标生物的检测水平;然后,优化损失函数,提高训练过程中的收敛速度和回归精度;最后,针对算法对部分类别检测效果较差的问题,在检测流程末端添加正负样本检测环节,降低误检率。实验结果表明,改进后的M＿YOLOv5s在测试集上的平均精度均值为82.4%,精确率为86.5%,召回率75.8%,与YOLOv5s相比精度略有下降,但是参数量减少了2.16MB,计算量降低了42.1%,实现了精度与速度的平衡。（3）针对现有算法在追踪水下生物时存在精度较差、目标易丢失的问题,基于EfficientNet网络中的MBConv模块重新设计DeepSORT算法的重识别网络,增强算法的特征提取能力。然后设计了水下运动目标识别与追踪系统,利用该系统评估改进后的水下图像增强和目标识别算法,并将运动生物的轨迹可视化。本文为解决复杂水下环境下的目标检测与轨迹追踪提供了思路。实验结果表明,基于深度学习的水下图像增强、目标检测和轨迹追踪算法效果较好,具备一定的应用价值。更多还原

【Abstract】 As an important field of marine scientific research,vision-based underwater target recognition and tracking technology can monitor the distribution and behavior of marine organisms in real time,which is of great significance to the protection of marine environment and ecology.However,the complex and changeable underwater environment brings difficulties and challenges to the identification and tracking of marine organisms.The original underwater target image has serious blue-green color cast,low contrast and blurred details.Meanwhile,the existing algorithms have problems such as low recognition efficiency and easy loss of tracking targets.Therefore,this thesis launched the research on underwater image enhancement,target recognition and tracking algorithm based on deep learning.The specific content is as follows:（1）Aiming at the problems of blue-green color cast,low contrast and blurred details in underwater images,an underwater image enhancement algorithm based on UWNET network is proposed.First of all,due to the lack of high-quality paired datasets,the CycleGAN（Cycle-Consistent Generative Adversarial Networks）network is used to render the aerial image into an underwater style image,which is used as the input of the underwater image enhancement network;secondly,I design a lightweight underwater image enhancement network UWNET.This model extracts the general features of underwater images through feature extraction networks,and uses the attention mechanism to enhance differentially the target information and background information in underwater images;finally,based on L₁ absolute error and structural similarity design loss function.The experimental results show that UWNET solves the blue-green color cast problem by minimizing the loss between the underwater image and the target image,improves the contrast,enriches the number of feature points,and has better comprehensive performance compared with the existing algorithms.（2）Aiming at the low efficiency of existing algorithms in identifying underwater organisms,a lightweight target detection algorithm M＿YOLOv5s is proposed based on YOLOv5s based network.Firstly,the backbone network is designed by using depth separable convolution and SE（Squeeze and Excitation）attention mechanism module to extract the features of underwater images;secondly,I use the Bi FPN module with stronger feature fusion ability to fuse the semantic information in the deep feature map and the texture information in the shallow feature map,which can improve the detection level of small target organisms;then,I optimize the loss function to improve the convergence speed and regression accuracy in the training process;finally,in view of the problem that the algorithm has a poor detection effect on some categories,a positive and negative sample detection link is added at the end of the detection process to reduce the false detection rate.The experimental results show that the average precision of the improved M＿YOLOv5s on the test set is 82.4%,the precision rate is86.5%,and the recall rate is 75.8%.Compared with YOLOv5s,the accuracy is slightly reduced,but the parameter amount is reduced by 2.16MB,and the calculation amount is reduced by42.1%.The improved model achieves a balance between accuracy and speed.（3）In view of the problems of poor accuracy and easy loss of targets in the existing algorithms when tracking underwater organisms,the re-identification network of the DeepSORT algorithm is redesigned based on the MBConv module of the EfficientNet network,in order to enhance the feature extraction ability of the algorithm.Then,an underwater moving target recognition and tracking system is designed.The improved underwater image enhancement and target recognition algorithms can be evaluated,and the trajectory of moving creatures is visualized by the system.This thesis provides ideas for solving target detection and trajectory tracking in complex underwater environments.The experimental results show that the underwater image enhancement,target detection and trajectory tracking algorithms based on deep learning are effective and have certain application value.更多还原