Combining Local and Global Features for Quality Assessment of Stitched Images in Virtual Reality

High-quality stitched images, namely wide-angle images stitched from small viewpoint images using stitching algorithm, are a key component for immersive VR experience. There is no stitching algorithm that can achieve perfect stitching for all visual scenes. To design better stitching algorithms, an accurate quality metric for stitched panoramic images is desired. In this paper, we propose a new quality assessment metric that focuses on both global and local distortions in stitched images. Specifically, we have performed the positioning of the distorted area and divided the panoramic image into distorted area and non-distorted area. For local quality, we separately extract the quality features of the distorted and non-distorted regions of the stitched image, and then calculate the distance between the features as a local quality metric of the stitched image. For global quality, we use general image quality evaluation features. The experimental results show that the combination of local and global features delivers significant performance improvement compared to the traditional image quality metrics.