VIGCN: an isotropic natural image stitching network based on graph convolution

Natural scene images contain a large number of repetitive structures and irregular contours, which increases the difficulty of stitching. Traditional image stitching methods are highly dependent on feature extraction and matching results, resulting in a large number of matching errors when stitching natural images. In recent years, some scholars have proposed several deep learning-based image stitching networks. However, the structure of the network has a great influence on the stitching results. Due to the limitation of the receptive field of the convolution kernel, the deep feature extraction process often requires the stacking of multiple convolutional layers, resulting in redundant network structures. To address this issue, we propose an isotropic network architecture (VIGCN), which segments the image into labels and transforms the graph structure through patch embedding, and then utilizes the GCN structure to capture global correlations instead of deep convolutional layers. We combine supervised and unsupervised loss functions to train the network on real natural image datasets, and the results show that the model exhibits good performance in both visible and infrared image domains.