An innovative transfer learning-polyp detection from wireless capsule endoscopy videos with optimal key frame selection and depth estimation

Abstract

In modern technology, the Wireless Capsule Endoscopy (WCE) is used to analyze and diagnose the small intestine part of the human body in a non-invasive manner. It consists of similar and non-modified information regarding texture and color, the absence of shot boundary makes the conventional keyframe mining and shot detection technique ineffective for this purpose. To resolve this issue, a deep learning-oriented keyframe mining method for extracting the keyframes from the captured video using the WCE procedure is suggested in this research work. With the help of benchmark databases, the endoscopic videos are collected. The gathered video is composed of numerous images frame in order to estimate the depth of the video. Here, the transfer learning techniques are adopted here for estimating the depths. The MobileNetV2 model’s encoder-decoder unit is attached to the UNet model of TransUNet + system to improve the rate of estimation of the depths. The deployed transfer learning model is named as TransUnet + with MobileNetv2 (TU-MNetv2). A new heuristic algorithm called the Improved Fitness-based American Zebra Optimization Algorithm (IF-AZOA) is implemented to select the ideal keyframes in terms of constraints like entropy information, moment of the images in depth estimated frame, key points and edge density. The estimated depth results are compared with the results obtained from several conventional classifiers and heuristic algorithms in order to prove the better performance obtained by the implemented depth estimation technique.