Treffer: Semantic segmentation by deep convolutional network

Ph.D. ; Object segmentation is a fundamental and long-standing problem in computer vision. It is defined as a multi-label classification problem, aiming to part ition different objects into different segments. According to the different types of inputs and goals, this task can be further divided into sub-tasks, i.e., semantic image segmentation , video object segmentation and so on . This thesis mainly focuses on addressing the segmentation t ask by using deep learning techniques, include increasing accuracy and accelerating the algorithm. By studying two typical segmentation t asks, we propose three novel approaches to address the common challenges in object segmentation. ; To modeling contextual information in natural images, Markov Random Field (MRF) is widely used in t he semantic segmentation task. On the one hand, convolut ional Neural Network(CNN) has shown its pot ential for learning strong features and greatly advanced the field of segmentation. From these, we propose a novel framework to combine CNN and MRF to achieve the benefit of both, namely Deep Parsing Network (DPN). Specifically, DPN extends a contemporary CNN architecture to model unary terms and additional layers are carefully devised to approximate the mean field algorithm (MF) for pairwise terms DPN is able to achieve high performance by approximating one iteration of MF and makes MF easier to be parallelized and speeded up in GPU. Our method achieves state-of-the-art accuracy 77.5% on the PASCAL VOC 2012 dataset. ; Video segmentation is also a valuable topic and more applicable in the real-world application. However, conventional video segmentation methods oft en rely on temporal cont inuity to propagate masks and suffer from issues like drift ing and inability to handle occlusions and large displacement. We formulate an effective mechanism to prevent the target from being lost via adaptive object re-identification. Specifically, we proposed a novel DyeNet which joint the mask propagation module and reidentification module into a single ...