Deep Learning-Based Image Stitching Technique Using Transformer
- Resource Type
- Article
- Authors
- 최병찬; 신승원; 김지현; 안세환; 주지한; 남해운
- Source
- 한국통신학회논문지, 48(12), pp.1577-1580 Dec, 2023
- Subject
- 전자/정보통신공학
- Language
- English
- ISSN
- 2287-3880
1226-4717
Image stitching is an image processing technique that combines multiple overlapping images into one. Classical image stitching extracts common feature keypoints from input images and use them as reference points for aligning the images. Convolutional neural network-based image stitching network combines latent feature vectors to produce stitched output from input overlapping images. This paper proposes a transformer-based image stitching network along with its implementation and training strategies. Unlike previous methods, the proposed transformerbased image stitching network explicitly learns to produce stitched output images from connection between patches of input images. It achieves 3.7dB higher PSNR and 0.12 higher SSIM than classical image stitching technique.