Edge Detection Cascaded with Simplified WDSR of IR Super Resolution

Edge Detection has been one of the techniques that can be used to solve super resolution problems. In this work, we present a research that was based on the WDSR (Wide Activation Super Resolution) model and the combination of three edge detection techniques: Sobel, Kirsch, and Prewitt. For the model portion, we simplified the original WDSR model by reducing the number of convolutional layers in each residual block and decreasing the number of residual blocks to cut down the computational burden of our computer. As for the dataset, we used our own high-resolution IR dataset that contained 7968 IR images for training and 1359 images for validation. These data are used to perform bicubic interpolation to create low-resolution images and the low-resolution images were paired with the original data to observe the result. By the end of our work, we managed to show that the model was capable to successfully reduce the computational time by approximately 11% from the baseline while maintaining the quality of super resolution for IR image within roughly 1% difference.