OmiQnet: Multiscale feature aggregation convolutional neural network for omnidirectional image assessment

Recently, deep learning-based methods for quality assessment of omnidirectional images (OIs) have gained widespread attention. However, existing methods face challenges because most omnidirectional image quality assessment (OIQA) methods inadequately consider projection distortions and visual complexity. In response, a multiscale feature aggregation convolutional neural network is proposed for OIQA to explore the feasibility of using multiscale features to strengthen the perception of projection distortion information. Specifically, cubemap projection (CMP) is employed to generate viewport images from equirectangular projection (ERP) images to effectively preserve more omnidirectional information. Subsequently, a multiscale feature extraction (MFE) module is designed to extract features at different levels and enhance the representation of distortion information. Additionally, a feature aggregation (FA) module is introduced to fuse multiscale features and fully improve the interconnection capability of the network. Finally, a quality regression (QR) module is employed to map the features to a quality score. Extensive experiments demonstrate the effectiveness and superiority of the proposed network over other state-of-the-art methods for accurately assessing OI quality.