cimese_net 0.37

Basic implementation of a Siamese network structure for detecting copyright infringement

A neural network is implemented here to detect copyrigt infringement violations in newly uploaded video files. The model takes the general form of a Siamese network, with two images filtered through the same convolutional neural network before a classification is made concerning the probability that the two images are a match. In order to train the model, frames were extracted from a set of high-resolution video files of movies and a corresponding set of lower-quality recordings of those movies. Randomly selected frames from the high-quality files were paired with the corresponding frame from the recorded version (a match) as well as a frame from another movie (not a match); these are referred to as triplets. Each of the images is run through the initial set of convolutional layers, which take the structure and weightings from the pre-trained VGG16 neural network, and a vector of length 4096 is returned. This serves as the input to a new set of top layers meant to classify the paired images as a match or not. In addition to the feature extraction and classification model, a means of aligning the recorded clip with the full movie is implemented to optimize the neural net performance. The top level function outputs a single probability of infringement that emerges as the average of the predicted match probabilities along the length of the potentially infringing clip.