结合RefineNet的无监督图像拼接网络  

Unsupervised Image Stitching Network Combined with RefineNet

作　　者：齐梦妍 朱磊[1] 万娅娅 QI Mengyan;ZHU Lei;WAN Yaya(Xi'an Polytechnic University,School of Electronics and Information,Xi'an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《长江信息通信》2022年第11期23-26,共4页Changjiang Information & Communications

摘　　要：为减少无监督图像拼接结果中存在的伪影和错位问题,提出了一种结合RefineNet的无监督图像拼接网络。该网络首先通过一个深度单应性(Deep Homography)算法计算两幅图像的投影矩阵,根据矩阵参数扭曲其中一副图像完成拼接过程的预对齐;其次在进行编解码处理时充分利用不同尺度的中间层特征图提取空间和语义信息,以加强网络对图像错位区域的感知能力;最后在恢复图像时引入链式残差池化结构捕获图像背景区域中的上下文特征,进一步提高了拼接的质量。实验结果表明:本文方法没有明显的拼接痕迹,亮度和颜色过渡自然,同时有效减少伪影和错位现象,结构相似性相较于其它3种拼接算法性能至少提高了1%,峰值信噪比较深度学习方法提高了0.26dB。To reduce the artifacts and misalignment problems in unsupervised image stitching results, an unsupervised image stitching network combined with RefineNet is proposed. The network first calculates the projection matrix of the two images through a Deep Homography algorithm, and distorts one of the images according to the matrix parameters to complete the prealignment of the splicing process;secondly, it makes full use of the middle of different scales during encoding and decoding.The layer feature map extracts spatial and semantic information to enhance the network’s ability to perceive image misplaced areas;finally, a Chain Residual Pooling structure is introduced to capture the contextual features in the background area of the image when restoring the image, which further improves the quality of stitching. The experimental results show that the method in this paper has no obvious splicing traces, the brightness and color transition are natural, and the artifacts and dislocations are effectively reduced. Compared with the other three splicing algorithms, the performance of the structural similarity is improved by at least 1%. The deep learning method improved by 0.26dB.

关 键 词：图像拼接 无监督网络 RefineNet 链式残差池化 HOMOGRAPHY 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]