Introduction to the special section “pattern recognition for Recent and Future Advances On intelligent systems” (SS:ISPR22)

Stereo image super-resolution (StereoISR) aims at recovering high resolution (HR) details and realistic textures from low-resolution (LR) left-right pairs with stereo correspondence. Recently, some advanced 2D methods have achieved good performance in realistic texture reconstruction via texture attention mechanism in a reference-based training way. However, the existing 3D super-resolution (SR) datasets lack the HR reference to obtain the LR-to-HR reference pair for the reference-based training. To overcome this problem, we propose a novel StereoISR network with a joint texture and parallax attention module (JTPAM), where the down-sampled LR (named super-low-resolution (SLR)) image and the LR image are taken as the SLR-to-LR reference pair, to simulate the texture conversion relationship between the LR and HR images. Additionally, the parallax attention mechanism is utilized to maintain the stereo correspondence by calculating the feature similarity along the Epipolar line. Experiment results on the Flickr1024, Middlebury, KITTI 2012 and KITTI 2015 datasets show that our network achieves the state-of-the-art (SOTA) performance over the compared methods, and the stereo SR images generated by our network have abundant binocular information and clear texture details.