... that conventional pixel-level reconstruction cannot effectively extract the global semantic information of the image, a novel model with the combination of attention mechanism and self-supervised learning method is proposed. First, an auxiliary task, reconstruct rotated image, is set to enable the network to learn...