Novel Fuzzy Entropy Based Leaky Shufflenet Content Based Video Retrival System

Abstract

Developing Content-Based Video Retrieval (CBVR) for large-scale videos is a major challenge due to the enormous growth of video content on the internet. One of the major downfalls associated with CBVR is high search response time and low accuracy. In this paper, a Novel Fuzzy entropy based Leaky ShuffleNet CBVR system has been proposed to reduce the search response time and for high accuracy. The proposed system has three major phases i) Data Processing, ii) Feature Extraction and iii) Feature selection and iv) Similarity search. Initially, the video is processed using Apache storm to convert video into keyframes. Consequently, facial landmarks, head pose, and eye gaze, edges are extracted using various feature extraction techniques. The most relevant features are selected in the feature selection phase by using the Fuzzy entropy measure. Finally, based on the selected features Leaky ShuffleNet retrieve the relevant videos based on the user query. Experiments were performed on two different datasets such as Hollywood2 and UCF50 in three different setups (Single Node, Vertical Scaling, Horizontal Scaling). Several metrics were analyzed to measure the effectiveness of the suggested strategy, including recall, specificity, accuracy, precision, and the F-measure. According to experimental results, the proposed system has a search reaction time of 0.75 seconds, which is lower than the existing methods. The proposed method improves the overall accuracy by 1.2%, 2.5%, and 3.2% better than the existing ECBVR-ACNN, FALKON, and EE-CBVR respectively.