Due to the widespread application of the virtual reality (VR) technique, omnidirectional image (OI) has attracted remarkable attention both from academia and industry. In contrast to a natural 2-D image, an OI contains $360^{\circ } \times 180^{\circ }$ panoramic content, which presents great challenges for no-reference quality assessment. In this article, we propose a saliency-guided no-reference OI quality assessment (OIQA) method based on scene content understanding. Inspired by the fact that humans use hierarchical representations to grade images, we extract multiscale features from each projected viewport. Then, we integrate the texture removal and background detection techniques to obtain the corresponding saliency map of each viewport, which is subsequently utilized to guide the multiscale feature fusion from the low-level feature to the high-level one. Furthermore, motivated by the human way of understanding content, we leverage a self-attention-based Transformer to build nonlocal mutual dependencies to perceive the variations of distortion and scene in each viewport. Moreover, we also propose a content perception hypernetwork to adaptively return weights and biases for quality regressor, which is conducive to understanding the scene content and learning the perception rule for the quality assessment procedure. Comprehensive experiments validate that the proposed method can achieve competitive performances on two available databases. The code is publicly available at https://github.com/ldyorchid/SCP-OIQA.