Comparative study on super resolution techniques for upper gastrointestinal endoscopic images

Endoscopy is considered the gold standard for diagnosis of gastrointestinal diseases. Image quality is an important creteria for a better accurate prediction of the diseases. Actually, in many current health facilities in developing countries as Vietnam, due to the endoscope limitation and environmental impacts, endoscopic images are of very low resolution. As a result, some textures and colors in lesion regions of the image could be ignored. This paper investigates different techniques for enchancement of image resolution. Spefically, we implement fundamental interpolation methods such as Nearest Neighbor Interpolation (NNI), Bilinear Interpolation (BLI) and Bicubic Interpolation (BCI) and advanced methods using deep learning such as Efficient Supixel Convolution Neuron Network (ESPCN), Residual Dense Network (RDN) and Super Resolution Dense Network All (SRDenseNet All). We then compare the performance of these techninques according to SSIM, PSNR and framerate metrics. The experimental results on dataset of upper gastrointestinal endoscopic images show that deep learning super-resolution method (RDN) provides the highest efficiency. This method produces sharper images, some of them look more intuitive and provide more information to doctors that can improve their diagnosis and treatment.