FALKON: Large-Scale Content-Based Video Retrieval Utilizing Deep-Features and Distributed In-memory Computing

In the boom of the big-data era, social media, 5g, High-speed Internet, and High-Tech Vision equipment have resulted in an enormous amount of video data production at an alarming speed. A fast and efficient content-based video retrieval system is not only desirable but essential in many domains with widespread applications. Conventional video retrieval techniques are inadequate to fulfill the needs of the day and keep pace with the rate of video production due to three core challenges: the sheer volume of the videos being produced, the complexity of video data, and redundancy in the video data. Due to these challenges, video operations & processing demand an enormous computing power and resources to process effectively. In this paper, we propose FALKON, a content-based video retrieval system harnessing the power of big-data technologies, deep-learning and distributed in-memory computation. First, we perform structural analysis on the videos, then spatial and temporal features are computed and indexed. We apply various optimization techniques to accelerate video processing as well as accuracy. We introduce VidRDD as a basic unit for distributed in-memory video computation. Furthermore, we introduce Video Query Maps as a relevance feedback mechanism to make the proposed system more reliable, user-friendly, and to improve the retrieval results. We implement FALKON on Hadoop, Hbase, Spark, and OpenCV. We achieve an average accuracy of 97.3%. Our evaluation results show that FALKON performs very well in terms of efficiency, scalability, computation time, and precision.