报告时间：2016年5月20日下午1点半

报告地点：软件学院综合楼5楼第一会议室

报告人：哈尔滨工业大学左孟旺教授

报告人题目：Discriminative Learning in Image Restoration

摘要：In this talk, we summarize our work on image restoration from three aspects. First, given a test noisy image, we described an approach to estimate hyper-parameters of gradient-based image model, and suggested a gradient histogram preservation method for image denoising. Second, give a set of image pairs of blurring/clean images, we present a discriminative learning method to learn iteration-wise priors for blind deconvolution. Iteration-wise loss, constraint on solution paths, and multi-scale scheme are also incorporated to guarantee its success. By introducing the concept of dynamic guidance, we provide a model for learning weighted analysis representation for guided depth image super-resolution. Finally, we investigate the possiblity to design CNN guided by image modeling, and learn a plain CNN model for image denoising.

报告人简介：左旺孟，哈尔滨工业大学计算机学院教授、博士生导师。IEEE会员，中国计算机学会会员，国际期刊《ISRN Signal Processing》编委。主要从事图像增强与复原、稀疏表达和深度学习等方面的研究。在ACM CSUR、IEEE TIP等重要国际期刊和CVPR、ICCV、ICPR、ICIP等重要国际会议上发表学术论文40余篇。曾获得全国优秀博士论文提名奖、哈尔滨工业大学优 秀博士论文奖和国际会议优秀论文奖。