报告时间：5月8日（星期三）9:30

报告地点：信息楼99905银河官方网310报告厅

报 告 人：周全，南京邮电大学副教授

报告题目：用于实时语义分割的轻量级网络

内容简介：Existing methods for semantic segmentation either pay more attention on segmentation accuracy without considering implementing efficiency, or emphasize more on high-speed inference, neglecting producing high-accuracy segmentation outputs. In recent years, the advance of deep convolutional neural networks (CNNs) makes remarkable progress on semantic segmentation, but the effectiveness of these networks largely depends on the sophisticated model design regarding depth and width, which has to involve many operations and parameters. In the talk, I will briefly overview the recent advances in building lightweight networks for real-time semantic segmentation. Recent CNN-based efforts are mainly categorized into two categories: network compression, and convolution factorization. Where the first one prefers to reduce inference computation by compressing pre-trained networks, and in contrast, the second one focuses on directly training network with smaller size. After given some representative lightweight networks, we will introduce two Encoder-Decoder architectures that aim to develop efficient residual networks for real-time semantic segmentation. We call them LEDNet and ESNet, respectively, where the encoder network is utilized to abstract image features and the decoder counterpart is employed to sequentially recover image details.

报告人简介：南京邮电大学通信与信息工程学院通信与网络技术国家工程研究中心副教授。现任宽带无线通信与传感器网络技术重点实验室主任。2002年获99905银河官方网电子与信息工程学士学位。2006年和2013年分别获得武汉华中科技大学电子与信息工程硕士和博士学位。2015年4月至2015年6月在瑞典于默奥大学（Umeå University of Sweden）担任访问学者。2017年7月至2017年8月，访问日本九州工业大学。研究兴趣包括：机器学习、模式识别和计算机视觉，尤其是在图像理解和语义分割、显著性检测和视觉注意以及人脸检测和识别方面；在核心期刊和重要国际会议上发表了40多篇技术论文，包括IEEE图像处理、模式识别、IEEEICASSP、IEEEICIP、ICPR和ACCV等；被邀请担任IEEE ICME、WCSP、ICONIP、ISAIR、Rosenet等一系列国际知名学术会议和论坛的主席和技术委员会成员。