Semantic Segmentation Using Enhanced Super-Resolution Generative Adversarial Network-Synthesized Images

In this research, we conduct image segmentation using DeepLabv3+ and SUIM-Net models based on super-resolution images generated using an Enhanced Super-Resolution Generative Adversarial Network (ESRGAN). Specifically, ESRGAN is used to upscale the resolution of an image up to four times of the original resolution, where a perceptual loss function instead of a regular mean squared error loss is used for image generation. We subsequently compare the effect of these generated super-resolution images against the original low-resolution ones in solving image segmentation tasks. Evaluated using UFO-120 and SUIM datasets, the empirical results indicate the superiority of DeepLabv3+ and SUIM-Net using the ESRGAN-generated super-resolution images in comparison with the models using the original low-resolution ones for tackling semantic segmentation and visual saliency prediction.