Motion-Aware Deep Feature-Based Scalable Video Summarization for Wireless Capsule Endoscopy Videos

Wireless Capsule Endoscopy (WCE) is a highly effective and non-invasive diagnostic procedure that is specifically designed to detect any possible abnormalities present in the gastrointestinal (GI) tract. One of the key challenges gastroenterologists face is analyzing the long videos generated during this procedure to identify any diseases or abnormalities. Video summarization is proposed as a solution to address the issue of the long reviewing time of the WCE video. This paper presents a scalable and generic approach for Wireless Capsule Endoscopy Video Summarization, focusing on two crucial features—motion and deep learning. Motion-ware features are calculated using the Gunnar Farneback algorithm’s dense optical flow calculation technique, while the computationally light MobileNetV2 model is used to compute deep features. SVM performs the final classification. Moreover, this technique uses all predicted classes, which reduces the risk of missing critical information while generating the final video summary. The proposed methodology has been assessed against Precision, Recall, F1-score, and Accuracy metrics. It achieved a remarkable accuracy of 95%. The video summary generated by the proposed method preserves the temporal relationship of the frames and significantly reduces the video reviewing time.