COMPUSR: Computationally Efficient Unsupervised Super-Resolution Approach for Wireless Capsule Endoscopy

Abstract

Wireless Capsule Endoscopy (WCE) is an imaging technology for diseases related to Gastrointestinal (GI) track. However, due to limited hardware, the spatial resolution of acquired images is usually coarser leading to poor diagnostic quality. To enhance high-frequency details associated to edges, object boundaries, corners, etc., which are required for detection and recognition tasks, High-Resolution (HR) images are usually constructed through software-driven Super-Resolution (SR) techniques. Deep learning-based SR approaches has been frequently explored in the computer-vision community recently to increase the SR picture’s quality. However, the prime barrier associated with the presently available deep learning SR approaches is the usage of supervised training where, LR images are simulated by using known deterioration (for instance, bicubic downsampling) from the acquired HR images. It is apparent that the deep models trained on such LR-HR pair exhibits poor generalization on the real LR observation. On the other hand, it is challenging to collect large clinical datasets consisting of true LR-HR pair for supervised learning. To circumvent this problem, we propose a computationally efficient unsupervised approach for SR task using Generative Adversarial Network (GAN) for WCE images (referred as COMPUSR). The proposed model avoids LR-HR pair and is trained using unsupervised strategy to learn the degradation of LR observation. It is validated on a derivative dataset of the original Kvasir dataset consisting of 10, 000 training samples and reveals considerable improvement over the other state-of-the-art methods on subjective and quantitative evaluations. Remarkably, the proposed approach needs significantly less number of learnable parameters (i.e., 2.4M) and a number of GFLOPS (i.e., 91.56G) than that of other approaches.