Distribution Learning for Video Segmentation Applications

Abstract

With the increase in the number of deep learning networks, many excellent methods have been proposed for video segmentation tasks. However, most of the these methods are for learning pattern information. Not as much work has been done in the area of distribution information, which is also useful for video segmentation. Therefore, this work focuses on learning statistical distributions via neural networks for video segmentation tasks including background subtraction, vessel segmentation and crack detection. In this thesis, we discuss four proposed methods in order to identify an effective way to learn statistical distributions. First, we propose a dynamic deep pixel distribution learning (D-DPDL) method for background subtraction. In D-DPDL, a random permutation of temporal pixels feature is used to force the network to learn the statistical distributions. Compared with previous background subtraction methods based on deep learning networks, the D-DPDL model only requires limited ground-truth frames for training, and it is effective even when training videos and testing videos are captured from different scenes. Then, we improve the D-DPDL method and apply it to vessel segmentation and crack detection, and we found that a wide rather than deep network works better. Finally, we proposed an arithmetic distribution neural network (ADNNet), which is based on arithmetic distribution layers, for learning distributions. The arithmetic distribution layers is the first work to propose network layers based on arithmetic distribution operations, which perform even better than convolutional layers in distribution classification.