| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 107 === Recent research on super resolution has progressed with the development of deep convolutional neural networks (DCNNs). Despite the use of faster and deeper convolutional neural networks, breakthroughs have been made in the accuracy and speed of single image super resolution. On the other hand, how do we restore finer texture details? However, we have artifacts appearing in the large upscale image of super resolution. In order to further improve the visual quality and the optimized super resolution method, the goal is mainly achieved by the selection of the loss function. In this thesis, we have developed a deep residual network that can reuse the previous feature maps. Our purposed model is constructed by the combination of two architectures, ResNet and DenseNet. We can adjust the size of the training image and the batch size during the training. According to this, we can elaborate on them that affect the performance of the deep residual network, say PSNR. Benefiting from these improvements, our deep residual network is able to recover photo-realistic textures from multiple downsampled images in benchmarks, which outperforms the methods such as SRCNN, VDSR, LapSRN, and SRResNet. The experimental results reveal that our developed deep residual network is also better than the NTIRE2017 champion winner EDSR does.
|
|---|
