一种改进的Mask R-CNN的图像实例分割算法

An Improved Mask R-CNN Image Instance Segmentation Algorithm

实例分割需要考虑像素级的分类准确性和目标实例级的高级语义特征,这是非常具有挑战性的。本文对Mask R-CNN算法进行改进,引入混合空洞卷积思想,设计了一种新的Resnet-HDC特征提取网络,在扩大特征提取感受野的同时减少了图像信息的丢失,进一步提升掩膜预测的精度;在特征金字塔网络基础上增加自底而上的侧边连接网络,为底层特征的传递提供了新的传播路径,解决了特征金字塔网络高层特征不能有效包含底层几何信息的问题,提高了底层特征的信息利用率。在Cityspaces数据集上的测试结果表明,该方法有效地提高实例分割的精度。

To achieve instance segmentation,both the accuracy of pixel-level classifi cation and the high-level semantic features of the target image instance need to be taken in full consideration,which is quite challenging.In this study,the Mask R-CNN algorithm was improved and a new Resnet-HDC feature extraction network was designed by introducing the concept of Hybrid Atrous Convolution to expand the receptive field of feature extraction,reduce the loss of image information,and increase the accuracy of mask prediction.In addition,a bottom-up side connection network was added to the feature pyramid network to provide a new propagation path for the transmission of the bottom-level features.By solving the ineffi cient containment of bottom-level geometric information in the high-level features of the feature pyramid network,it also improved the utilization of underlying features.The test results on the Cityspaces dataset showed that this method eff ectively improves the accuracy of instance segmentation.