Blind Stitched Image Quality Evaluator Based on Joint Tensor and Edge Sparse Representation

The goal of image stitching task is to seamlessly merge multiple narrow field of view images with overlapping regions into a stitched panoramic image (SPI). However, despite the existence of various existing image stitching methods, there is still a lack of in-depth research on how to effectively evaluate the quality of SPIs. To address this issue, a novel blind quality evaluator for SPIs based on tensor and edge sparse representation is proposed. Specifically, considering that image stitching disrupts edge structures as well as brightness and color information, two over-complete dictionaries in the edge space and tensor principal component space are first trained respectively. These dictionaries are defined as the edge dictionary and tensor dictionary. Given a test SPI, we initially extract its edge map and tensor principal component map. Then, with the learned two over-complete dictionaries, sparse coding is performed on each block of both the SPI’s edge map and tensor principal component map to generate the corresponding sparse coefficients. Additionally, by analyzing statistical distribution characteristics of these coefficients, a log-normal pooling scheme for absolute values of sparse coefficients is designed to enhance feature aggregation effectiveness. Finally, the extracted perceptual sparse features are concatenated and pooled into an objective quality score through regression function. Experimental results demonstrate that compared to classical 2D image quality metrics and SPI quality metrics, our proposed approach can more accurately assess unique distortions caused by image stitching while maintaining better consistency with subjective ratings.