An Implementation Of A Rate Controller Using A Neural Network

The rate controller is a crucial component of a video encoder. The component regulates bit rate in a video, determining how much information is to be updated in a video frame by assigning an overall compression rate to said frame. At Axis Communications, a company specializing in network cameras, the component is implemented using traditional automatic control methods operating on a select few data parameters. This thesis project aims to implement the rate controller with a machine learning approach, using a neural network. A key challenge is the task of devising a set of customized quality measurements suited for the purposes of surveillance video. Mainly, these measurements are based around object detection. The quality measurement deemed most promising for future use is based on the presence of false positives and negatives in object detection data. The measurements are used to generate a set of labels, each a theoretically optimal compression rate. The input features used by the network is video meta data, a set of parameters describing video content that is more low-dimensional than pixel data from the video. An attempt at encoding videos with a rate controller implemented using a set of trained neural networks is finally carried out. The results indicate that the networks can be trained to adapt compression rate based on changes in activity within a video. However, the compression rates given by the network can change drastically from frame-to-frame. Also, the ability of the networks to adapt to smaller changes can not be reliably determined, requiring further testing