Image super-resolution for endangered animals using refined super-resolution generative adversarial network

Abstract

In the nature and wildlife protection field, biologists want to learn about endangered animals' behavior, environment, and activities. The hope is to get more information about endangered animals, such as their diets, predators, and surroundings. With the help of digital devices, biologists use drones to capture information about the changing environment of endangered animals. Therefore, high-resolution images with precise details are essential in the research process. Recently, there have been many deep learning models for improving image resolution in the computer science field. For example, super-resolution convolutional neural networks (SRCNN), efficient sub-pixel convolutional neural network (ESPCN), deeply-recursive convolutional network (DRCN), and super-resolution generative adversarial network (SRGAN). SRGAN is a deep learning network model based on generative adversarial networks. It has a good effect on improving image resolution. Based on the SRGAN model, the paper refined the loss function and structure of the SRGAN networks, which optimized the details of the generated high-resolution images, and provided more accurate high-resolution image results. The refined SRGAN aims to give better results in animal image analysis. We selected DIV2K dataset as the training and testing data and compared the refined SRGAN method with the original SRGAN method. The quantitative evaluation used the peak signal-to-noise ratio (PNSR) and structural similarity index measurement (SSIM). Compared with the SRGAN model with one convolution layer, the model improves the average PSNR value of 0.93dB, and the SSIM value reaches 0.8723. With the help of the improved SRGAN network, biologists can obtain better super-resolution images under the premise that other conditions remain unchanged to protect wild animals.