卷积神经网络低层特征辅助的图像实例分割方法  

Low-level CNN Feature Aided Image Instance Segmentation

作　　者：樊玮[1] 刘挺 黄睿 郭青[2] 张宝[2] FAN Wei;LIU Ting;HUANG Rui;GUO Qing;ZHANG Bao(College of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China;College of Intelligence and Computing,Tianjin University,Tianjin 300350,China)

机构地区：[1]中国民航大学计算机科学与技术学院,天津300300 [2]天津大学智能与计算学部,天津300350

出　　处：《计算机科学》2020年第11期186-191,共6页Computer Science

基　　金：天津市教委科研计划项目(2019KJ126);中国民航大学中央高校基金项目(3122018C021,3122018C020)。

摘　　要：流行的实例分割网络Mask R-CNN在进行实例分割时,存在目标分割边界和分割轮廓粗糙的问题,导致分割精度低。针对此问题,提出在Mask R-CNN分割分支中引入网络的低层卷积特征进行高精度的实例分割方法。首先从特征提取网络中选择特征,通过插值算法将其缩放至固定尺度(输入图像的1/8)作为低层特征;然后通过RoI对齐操作提取当前待分割目标的特征后与原始的Mask R-CNN的分割分支对应目标的特征进行拼接,并将其作为精细化目标分割的特征。低层网络特征引入了更多的低级纹理和轮廓信息,可以有效地提高物体的分割精度。在COCO2017数据集上,所提方法使用ResNet-101-FPN作为特征提取网络得到的分割结果的平均准确度(AP)相对于Mask R-CNN提高了1.2%。实验结果表明,所提方法在使用不同特征提取网络时具有较好的鲁棒性和有效性。The popular instance segmentation network,Mask R-CNN,has rough target segmentation boundaries and segmentation contours when performing instance segmentation,which leads to low segmentation accuracy.To solve this problem,a high-precision instance segmentation method is proposed by introducing the low-level features of the network into the segmentation branch of Mask R-CNN.Specifically,it selects the convolutional features from lower layers of feature extraction network at first.And then,it resizes the features to a fixed scale(1/8 of the input image)by interpolation algorithm to form the low-level features.It concatenates the features of original segmentation branch of Mask R-CNN with the features extracted by RoI Align ope-ration from low-level features for current target.Since low-level features introduce more low-level texture and contour information,it can effectively improve the accuracy of instance segmentation.Compared with Mask R-CNN,the proposed method obtains 1.2%relative average precision(AP)improvement on the COCO2017 dataset by using ResNet-101-FPN as the feature extraction network.Experimental results show that the proposed method is robust and effective when using different feature extraction networks.

关 键 词：深度学习 深度神经网络 实例分割 特征融合 低层特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]