No-reference stereoscopic image quality assessment based on global and local content characteristics

No-reference stereoscopic images quality assessment (NR-SIQA) via deep learning has gained increasing attention. In this paper, we propose a no-reference stereoscopic image quality assessment method based on global and local content characteristics. The proposed method simulates the perception route of human visual system, and derives features from the fused view and single view through the global feature fusion sub-network and local feature enhancement sub-network. As for the fused view, a cross-fusion strategy is applied to model the process in the V1 visual cortex, and the multi-scales pooling (MSP) is utilized to integrate context information under different sub-regions for effective global feature extraction. As for the single view, the asymmetric convolution block (ACB) is introduced to strengthen the local information description. By jointly considering the fused view and single view, the proposed network can efficiently extract the features for quality assessment. Finally, a weighted average strategy is applied to estimate the visual quality of stereoscopic image. Experimental results on 3D quality databases demonstrate that the proposed network is superior to the state-of-the-art metrics, and achieves an excellent performance.