基于A-HRNet对抗自编码器的矩阵填充算法研究

作者	黄梓玉
姓名汉语拼音	HuangZiyu
学号	2021000003054
培养单位	兰州财经大学
电话	13766986158
电子邮件	zziyuh@163.com
入学年份	2021-9
学位类别	学术硕士
培养级别	硕士研究生
学科门类	经济学
一级学科名称	应用经济学
学科方向	统计学
学科代码	020208
第一导师姓名	黄恒君
第一导师姓名汉语拼音	HuangHengjun
第一导师单位	兰州财经大学统计与数据科学学院
第一导师职称	教授
题名	基于A-HRNet对抗自编码器的矩阵填充算法研究
英文题名	Research on matrix filling algorithm based on A-HRNet Adversarial Auto-encoder
关键词	深度学习矩阵填充注意力机制 HRNet 深度特征
外文关键词	Deep learning ; Matrix filling ; Attention mechanism ; HRNet ; Depth characteristics
摘要	矩阵填充技术提供了一种高效的方式来弥补数据的缺失。数据缺失是数据分析、机器学习、图像处理等诸多领域中的共性问题。利用矩阵填充，可发现隐藏在数据中的规律及变化规律，从而对其进行更深层次的认识和运用。矩阵填充还可以改善数据的完整性和质量，提高预测和决策的准确性。因此，在数据处理和实际应用中，解决矩阵填充的问题显得尤为关键。矩阵填充涉及多种数据类型，图像类填充是一个关键问题，因此构建可修复图像类数据的矩阵填充模型就显得尤为重要。通过填充缺失的像素，可使图像恢复完整，提高图像的可视化效果。图像填充还能优化图像的后续处理流程，例如图像的分类、对象的识别以及图像的重建工作。因此，在图像处理和分析领域，解决图像填充的问题具有不可忽视的重要性。在此基础上，提出一种融合 HRNet 与注意力机制的对抗式自编码模型 AH-AAE。AH-AAE 模型从两个方面对基本对抗自编码器进行了改进。第一个改进的点是将生成器中的编码器部分进行重构为 HRNet，使其能够更好地表达和保持图像的细节，进而改善修复结果的质量和逼真度。第二个改进的地方是将注意力机制引入生成器中，注意力机制通过图像细节加强了对局部填充的控制，提高图像填充质量。当图像部分受损时，可以保证填充的结果只覆盖需要的部分，而不会对整幅图像做多余的改动。本文提出的 AH-AAE 模型通过引入通道注意力在跳跃连接处构建通道相似性融合模块来丰富通道之间的特征关系；在解码器网络中，结合空间注意力和位置编码的位置融合模块用于增强边界位置信息的表达。为了验证模型效果，在 MS-COCO 以及 KITTI 数据集上进行了多项实验，证明了 AH-AAE 模型的填充能力、去噪能力以及环境感知能力，在多个指标上均达到了最优。本文所提的模型性能良好，未来可为敦煌壁画修复、计算机视觉以及视频填充等图像类数据修复领域发展提供参考。
英文摘要	Matrix filling technology provides an efficient way to make up for the lack of data. Data missing is a common problem in many fields such as data analysis, machine learning and image processing. By using matrix filling, we can find the laws hidden in the data and the changing laws, so as to understand and apply them in a deeper level. Matrix filling can also improve the integrity and quality of data and improve the accuracy of prediction and decision-making. Therefore, in data processing and practical application, it is particularly critical to solve the problem of matrix filling. Matrix filling involves many data types, and image class filling is a key problem, so it is particularly important to build a matrix filling model that can repair image class data. By filling the missing pixels, the image can be restored completely and the visualization effect of the image can be improved. Image filling can also optimize the subsequent processing flow of images, such as image classification, object recognition and image reconstruction. Therefore, in the field of image processing and analysis, it is of great importance to solve the problem of image filling. On this basis, an adversarial auto-encoding model AH-AAE integrating HRNet and attention mechanism was proposed. The AH-AAE model improves the basic adversarial auto-encoder from two aspects. The first aspect of the transformation is to reconstruct the encoder part of the generator into HRNet, which allows it to better express and preserve the details of the image, thereby improving the quality and fidelity of the restoration results. The second aspect of the transformation is the introduction of attention mechanisms into the generator, which enhances the control of local fill through image details, improving the quality of image fill. When part of an image is damaged, you can be sure that the result of the fill will cover only the part you need, without making unnecessary changes to the entire image. The AH-AAE model proposed in this paper enriches the feature relationship between channels by introducing channel attention and constructing a channel similarity fusion module at the hopping connection. In the decoder network, the position fusion module combining spatial attention and position coding is used to enhance the expression of boundary position information. In order to verify the effect of the model, a number of experiments were carried out on MS-COCO and KITTI datasets, which proved that the filling ability, denoising ability and environment perception ability of the AH-AAE model reached the optimal in multiple indicators. The model proposed in this paper has good performance and can provide a reference for the development of image data restoration fields such as Dunhuang mural restoration, computer vision, and video filling in the future.
学位类型	硕士
答辩日期	2024-05
学位授予地点	甘肃省兰州市
语种	中文
论文总页数	74
参考文献总数	80
馆藏号	0005655
保密级别	公开
中图分类号	C8/364
文献类型	学位论文
条目标识符	http://ir.lzufe.edu.cn/handle/39EH0E1M/36917
专题	统计与数据科学学院
推荐引用方式 GB/T 7714	黄梓玉. 基于A-HRNet对抗自编码器的矩阵填充算法研究[D]. 甘肃省兰州市. 兰州财经大学,2024.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
2021000003054.pdf（4364KB）	学位论文		暂不开放	CC BY-NC-SA	请求全文