Rapid Video Deduplication based on Global & Local Features using Convolution Neural Network

Video deduplication faces significant difficulties because of the large video data explosion. In recent years, deduplication algorithms have made encouraging strides, particularly with the aid of binary hashing techniques. However, research on the general hash-based architecture and the effective similarity ranking technique for video deduplication has been sparse up to this point. The existing hashing techniques finds duplication of videos in the form of binary hash at file level i.e., entire video (Global feature comparison). The proposed research focuses on developing a versatile and quick video deduplication methodology which checks duplication of videos at frame level (Local feature Comparison). If the input video is not entirely duplicated (global feature extraction), the video is then divided into frames to check duplication at frame level. To achieve this, the power of Deep Learning is exploited, using CNN, and a distinct compressed representation for each frame is produced, there are a little over 1000 features extracted for every frame and stored locally. The performance of the proposed methodology is measured using the Euclidean distance measure, that calculates the distance between the pre-trained data and the query video. The goal of the proposed local feature extraction is to assess the number of duplicate frames over the entire video. The proposed approach achieves an accuracy of 95.6% in finding out the number of duplicate frames and it has been concluded that it performs the best when compared with other state-of-art algorithms.