Cinematographic Shot Classification trough Deep Learning

The goal of this experimental thesis is to classify cinematographic shots by exploiting a convolutional neural network, also known as CNN. Usually CNNs are used to classify objects inside images, while in this case the CNN is used to classify images themselves using a partition similar to the frame partition typical of movie industry. The idea behind this project is to reduce the editing time of a video by classifying the video files. By doing so the video files will be divided according to the type of shot used. In this way during video editing it is known a priori where to look for a specific type of shots. The thesis is divided in 5 chapters. The first chapter starts with a quick introduction on machine learning, as a branch of weak AI. Then it digs deeper into the theory behind machine learning. After that there is an excursus on the different types of machine learning. The second chapter focuses on neural networks, a particular class of machine learning algorithms which is able to perform deep learning. The first part of the chapter explains how a neural network works once its training phase is done. The second part of this chapter is about how actually a neural network learns. The third chapter is about Convolutional Neural Networks, which is the type of neural network used in the thesis. The fourth chapter focuses on the work done for the thesis. The goal was to create a CNN able to classify images into cinematographic shots. The first part is a quick explanation on the different kinds of shots, while the second part focuses on the creation of the dataset used to train the neural network and on explaining why it was built the way it is. The third part of the chapter focuses on the specific CNN created for this thesis and on some methods to increase its accuracy. The fifth and final part of the thesis envisages possible future developments of the application with the author’s conclusions.