Video Retrieval Based on CNN Feature and Scalar Quantization

Abstract

In recent years, the video dissemination has become an important information medium with the development of the Internet and the rise of short video platforms, and infringements against long videos have followed, so an method of efficient and automated short video infringement detection is necessary. This paper proposes a method of video copyright detection based on CNN features and Scalar Quantizer, in which the deep convolutional neural network is used to obtain the decoded video frame’s feature vector, and then the Scalar Quantizer is used to search the feature vector based on the approximate nearest neighbor search, and finally the target video is determined by finding the shortest average Euclidean distance of the target video frames. This paper sets a distance threshold and a ratio threshold based on this method to form a new method, and then compares the recall and precision of two methods.