A deep learning-based image stitching method for fiber-optic image bundles imaging

For the problem of small field of view of fiber-optic image bundles imaging, an unsupervised fiber-optic image bundles imaging images stitching algorithm based on deep learning is proposed. The algorithm consists of two stages: unsupervised coarse image alignment and image reconstruction. In the first stage, we design a new backbone network that utilizes the Focus structure instead of down-sampling operations. We incorporate the ELAN structure and add attention mechanisms, which effectively enhances the backbone ’s accuracy in predicting the homography matrix while significantly reducing the model size. In the second stage, we used a more efficient convolution module C3 instead of regular convolution, reducing the number of model parameters and improving computational efficiency. Experiments are conducted using a dataset comprising fiber-optic image bundles imaging dataset and two publicly available stitching datasets (Warped MS-COCO and UDIS-D). The experimental results demonstrate that our method achieves high evaluation metrics and performs better in stitching real scene images with parallax and poor-quality fiber-optic image bundles. The network model is improved in terms of accuracy, size, and speed, meeting the practical requirements for fiber-optic image bundles imaging images stitching.