Endo-CV is an online image detection and segmentation challenge. The dataset used in the following code can be obtained from the https://endocv.grand-challenge.org/. The folder contains codes of various deep learning architectures and various preprocessing steps which were performed.

Endoscopy is a common non-invasive procedure used for both diagnosis and various minimal surgical procedures. The process involves a probe with a camera at the front which is inserted in the targeted body cavity. The procedure may require clear staticcamera images or even live video feed but, the camera feed gets routinely distorted. It is distorted by both physical factors like bubbles and debris along with software defects like pixel saturation and motion blur. Since these involve environmental factors, they cannot be merely solved by upgrading the hardware. This leaves room for the application of deep learning techniques, which give a probable look of the body cavity in the absence of defects. We propose a fully automatic framework that can: 1) detect and classify seven different primary defects, 2) provide a quality score for each frame. Existing state-of-the-art methods only deal with the detection of very domain-specific images. In an attempt to find the best method for our use-case we employ 3 different deep learning architectures and modify them according to our instance. In the end, we do a detailed analysis of these methods along with their pros and cons.