A novel image stitching method based on an unsupervised deep learning algorithm considering homography estimation and networks
Hao Hu, Ting Sun, Shijie Yu, Shuixin Deng, Baohua Chen · 2024
With the continuous development of digital photography technology and quality inspection, the demand for image stitching in practical applications is increasing. Traditional image stitching algorithms employ a variety of hand-designed methods for feature extraction, matching, and optimization. However, these traditional feature-based image stitching techniques heavily rely on feature extraction and may not perform well in scenarios with limited features. Current image stitching solutions based on supervised deep learning lack relevant data sets, and labeling data is relatively cumbersome, making supervised deep learning methods unreliable. At the same time, the rise of unsupervised deep learning algorithms provides new ideas for image stitching. We use unsupervised homography estimation to provide information about the geometric relationship between images, Stitching-Domain Transformer Layer to align feature maps, warp and generate masks, it helps to enhance the reality and continuity of splicing. We simultaneously utilize a pre-trained deep learning model (VGG) for feature extraction. We adjust the smoothness loss term to ensure smoother transitions within the stitching areas. Throughout the training process, we continuously optimize the number of convolutional layers, channels, and network depth to achieve optimal results. The superiority of the unsupervised learning algorithm compared to other classic algorithms was verified through experiments. Finally, we discussed the challenges and future applications of unsupervised deep learning in image stitching.