Video Copy Detection motivated by Image Classification using Sparse Coding

In the presence of a vast amount of digital video data, the protection of intellectual property of the creator is of utmost importance. Video copy detection, which is used to find copyright infringements, is also useful in video information retrieval. We propose a novel technique for video copy detection using an image classification framework and sparse coding. The underlying image classification framework is based on non-negative sparse coding, low-rank and sparse matrix decomposition techniques along with Spatial Pyramid Matching (LR-Sc+SPM). SIFT features from each image are encoded using non-negative sparse coding. Using Spatial Pyramid Matching + Max pooling, we capture the spatial relations between the sparse codes. Low-rank and sparse matrix decomposition is then used to exploit correlations and dissimilarities between images of the same class. Extending this to video copy detection, we create a framework where scene change detection is performed using edge maps. The resulting scenes are divided into classes and, using the same image classification framework, a multi-class linear SVM is trained. We evaluate our proposed algorithm against two state-of-the-art techniques for video copy detection with accuracy and computational time being the metrics.