【字色： 红 蓝 褐 绿 黑 紫 粉红 深蓝】
【字体:8 7 6 5 4 3 2 1】
题目：Robust Image labeling using Conditional Random Fields based Machine Learning
报告人：Xiao-Ping (Steven) Zhang
报告人简介：Xiao-Ping (Steven) Zhang received the B.S. and Ph.D. degrees from Tsinghua University, in 1992 and 1996, respectively, all in electronic engineering. He holds an MBA in Finance and Economics with Honors from the University of Chicago Booth School of Business. He is now Professor and Director of Communication and Signal Processing Applications Laboratory (CASPAL), with the Department of Electrical and Computer Engineering, Ryerson University. He has served as Program Director of Graduate Studies.
内容提要：Multiclass image labeling is a challenging task due to the limited discriminative power of low-level visual features in describing the diverse range of high-level visual semantics of objects. Dense conditional random fields (CRF) have obtained significant progress in labeling accuracy due to deployment of context information by imposing image-coherent label consistency and modeling object co-occurrence statistics. Dense random fields are confined to the success of the initial unary classifier including deep neural network (DNN) and very prone to over-smoothing small objects from “thing” classes in the large pool of pixels from image background. In this talk, we discuss two new CRF based machine learning models for robust image labeling. First, new feature functions based on generalized Gaussian mixture models (GGMM) are designed and their efficacy is investigated. This new model proves more successful than Gaussian and Laplacian mixture models. Second, we apply scene level contextual information to integrate global visual semantics of the image with pixel-wise dense inference of fully connected CRF to preserve small thing classes and to make dense inference robust to initial misclassifications of the unary classifier. Proposed inference algorithm factorizes the joint probability of labeling configuration and image scene type to obtain prediction update equations for labeling individual image pixels and also the overall scene type of the image.