Self-supervised Learning Aided Blind Stitched Panoramic Image Quality Assessment

For deep learning based stitched panoramic image quality assessment, it is costly to train a network model using a large-scale dataset with human annotation. Moreover, it is practical to evaluate image quality without reference images (i.e., Blind Stitched Image Quality Assessment, BSIQA). Selfsupervised learning (SSL) avoids the use of annotated large-scale training dataset while few BSIQA schemes of panoramic stitched images using SSL have been proposed. Thus this paper proposes a SSL aided BSIQA scheme for panoramas. The first training phase learns from a large-scale dataset, where SSL based image colorization is incorporated into supervised learning based classification for learning generalized visual representations. With transferred knowledge from the first phase, the second training phase learns from a small dataset with subjective scores for the task of BSIQA. Test results show that SSL indeed improves prediction accuracy of BSIQA of panoramas on ISIQA dataset.