Performance verification of deep image prior using two activation functions in super-resolution of food images

Abstract

In recent years, many people have posted cooking recipes and reviews on the Internet, while sometimes the image quality is poor and information cannot be transmitted well. To solve this problem, a method using CNN is effective, and one of them is a method of restoring resolution using only a single low-resolution image called deep image prior. This method uses a network consisting of a downsampling module, an upsampling module, and a skip module, and all modules include a convolutional layer and an activation function layer. In this paper, we compared the performance obtained by applying a different activation function to the skip module from other modules. The activation functions used are RReLU function, a ReLU function that has a random slope when the input value is negative, RReLU_1 and RReLU_2 that manipulate the probability of the random number. Then, the performance index of eight dishes images was compared by applying an evaluation index called PSNR. As a result of the experiment, it was found that the performance was improved by setting the different activation function of the skip module from that of the other modules.