多任务网络融合多层信息的目标定位-计算机辅助设计与图形学学报.PDF

第29 卷 第7 期 计算机辅助设计与图形学学报 Vol. 29 No.7 2017 年7 月 Journal of Computer-Aided Design Computer Graphics Jul. 2017 多任务网络融合多层信息的目标定位 1) 1) 1)* 1) 2) 田 彦 , 王慧燕 , 王 勋 , 黄 刚 , 章国峰 1) (浙江工商大学计算机与信息工程学院 杭州 310018) 2) (浙江大学CADCG 国家重点实验室 杭州 310058) (wx@zjgsu. edu. cn) : 目标定位是计算机视觉中的关键问题, 针对目前区域卷积神经网络及其扩展算法的精度受限于输出卷积图 尺寸、不能得到目标准确位置的问题, 提出一种基于多任务卷积神经网络的目标定位算法. 在特征提取阶段, 不同的 任务共享相同的特征提取层; 然后分别针对不同层次信息训练对应的后续网络; 在提取出高层和低层信息后, 融合 低层信息和高层信息得到准确的目标位置. 在PASCAL VOC 2007 数据库和交通场景数据库进行实验的结果表明, 该 算法可以有效地提高目标定位的精度. ：目标定位; 边缘信息; 多任务网络; 深层学习; 信息融合 ：TP391.41 Multi-Task Network Combing Multi-Level Information for Object Localization Tian Yan1), Wang Huiyan1), Wang Xun1)*, Huang Gang1), and Zhang Guofeng2) 1) (School of Computer Information Engineering, Zhejiang Gongshang University, Hangzhou 310018) 2) (State Key Laboratory of CADCG, Zhejiang University, Hangzhou 310058) Abstract: Object localization is a very important challenge in computer vision. At present, region convolutional neural networks (RCNN) and extensions had accuracies limited by the size of the output convolution map, hence they could not get exact object locations. In this paper, a new algorithm based on multi-task convolutional neural network was proposed for object localization, firstly different tasks shared feature extraction layer, and then the sub-networks for different levels of information were trained. When the high-level and low-level information were extracted, the accurate target location was obtained by a fusion of these two level information. Finally, the performance of the algorithm was verified on the PASCAL VOC 2007 database and the traffic scene database, and experiments show that the alg