Performance Comparison of Enhanced Deep Residual Networks for Single-image Super-Resolution Model with Change in Loss Function
Eunseo Lee, Taek Geun Whang-Bo · 2023
Super-resolution technology is receiving a lot of attention to restore low-resolution images to high-resolution. Among several models, we use an enhanced deep super-resolution network model that outperforms a single image. EDSR maximizes the efficiency of learning by using residual blocks with batch normalization removed from residual blocks of ResNet, enabling fast and good accuracy learning. We compared loss functions such as mean absolute error (MAE), mean square error (MSE), and binary cross-entropy (BCE) for resolution x4 bicubic down sampling images using Enhanced Deep Super-Resolution Network (EDSR). The mean absolute error (MAE) showed the best performance, and the mean square error (MSE) also showed compliance performance that could not be distinguished by the human eye. On the other hand, binary cross-entropy (BCE) showed significantly lower performance. In future studies, it is judged that a performance comparative experiment according to the optimal hyper-parameters for each scale is necessary. It is also an applicable technique for fast transmission of high-resolution images.