An Attention based Video Summarization Technique for Wireless Capsule Endoscopy Data

Wireless Capsule Endoscopy (WCE) data captured from a pill-sized camera typically results in an 8 to 12 hours video from the start of the digestive system to the end of it. Viewing and analyzing such lengthy videos is a tedious task when clinical practitioners have limited time availability. Viewing long videos can also induce fatigue making the practitioners miss key or important segments needed for determining pathology. Video Summarization (VS) techniques aim to provide a concise yet comprehensive summary by selecting the most relevant segments of the video. Summarized WCE videos can be very useful in clinical settings saving time and resources. In this paper, a new approach for VS on WCE data is proposed. The proposed VS approach employs Bidirectional Long Short- Term Memory (BiLSTM) with attention to WCE data while performing both forward and backward pass during training. To implement the proposed approach in a supervised manner, an annotation assignment was done on the Kvasir Capsule dataset. The proposed approach is validated empirically on the publicly available WCE video dataset of Kvasir Capsule videos where the average F-Score is noted as 84.57%.