Master of Science
In recent years, convolutional neural networks (CNNs) are leading the way in many computer vision problems. Since the development of fully convolutional networks, CNNs have been widely employed for low-level pixel-labeling problems, and successfully pushed the performance to a new level. Although CNNs are able to extract highly discriminative features, they typically assign a class label to each image pixel individually. This leads to various spatial inconsistencies. Therefore, CNNs are commonly combined with graphical models, such as conditional random fields (CRFs), to impose spatial coherence. CRFs were invented precisely for the task of imposing spatial coherence among image pixels. The coherence regularization weight serves an important role of controlling the regularization strength in the CRF optimization, and has a great influence on the quality of the final result. Traditionally this weight value is set to a fixed number for all images.
In this thesis, we propose a novel approach to learn the coherence regularization weight for each individual image using a CNN, and then apply this per-image-learned weight in the CNN+CRF system. We first construct a dataset where the optimal regularization weight for the CRF optimization has been pre-computed for each image. We adopt convolutional regression networks with standard architecture for learning, and tailor the input according to our problem. We test the effectiveness of our approach on the task of salient object segmentation where a graph-cut based CRF optimizer can generate globally optimal solution. We show that consistent performance improvements can be achieved by using the regularization weight learned on per-image basis as opposed to a fixed regularization weight for all images in the dataset.
Wu, Jiaxiao, "Learning Regularization Weight for CRF Optimization" (2018). Electronic Thesis and Dissertation Repository. 5414.