Self-supervised panoramic stitched image quality assessment based on contrastive learning

In recent years, with the rapid development of virtual reality technology and the advent of the 5G era, panoramic images have received increasingly widespread attention. Nowadays, researchers have proposed numerous image stitching algorithms. However, research on assessing the quality of stitched images is still relatively scarce. Furthermore, the stitching distortions introduced during the generation of panoramic content make the task of quality assessment even more challenging. In this paper, a new network for panoramic stitched image quality assessment is proposed. To be specific, this model contains two stages: the contrastive learning stage and the quality prediction stage. In the first stage, we introduce two pretext tasks as learning objectives: distortion type prediction and distortion level prediction. This allows the network to learn corresponding features from different viewpoints with varying distortion types and severities. During this process, we utilize prior knowledge of four pre-classified distortion types as category labels and three distortion severity levels as distortion severity labels to assist the pretext tasks. Subsequently, a universal convolutional neural network (CNN) model is trained using a pairwise comparison method. In the quality prediction stage, the trained CNN weights are frozen, and the learned feature representation is mapped to the final quality score through linear regression. We evaluate the proposed network on two benchmark databases and results demonstrate that the combination of two pretext tasks can obtain more accurate results. Overall, our method is superior to existing full-reference and no-reference models designed for 2D images and 360° panoramic stitched image quality assessment.