To solve the problem of low accuracy and low speed of linear flexible body segmentation,an improved DeepLabV3+network was proposed.The MobileNetV2,a lightweight and easy-to-converge feature extraction network,was utilized as the backbone network.The CA(channel attention)module was employed to concentrate on key information in the input features.A multi-scale atrous pyramid pooling structure was introduced to enhance the network's receptive field and training efficiency.In the decoding stage,an improved cascaded feature fusion module was proposed to fuse three shallow-level features,thereby improving the representation capability of semantic information.Experimental results show that compared with the non-improved network,the improved network improves MIOU and IOU by 2.82％and 3.46％,respectively,and reduces the prediction time by 5.2 ms,which can realize the accurate segmentation of deformable linear objects under complex background.
关键词
语义分割/轻量化网络/注意力机制/特征融合/线状柔性体分割/空洞卷积/级联特征融合
Key words
semantic segmentation/light-weight network/mechanism of attention/feature integration/deformable linear objects segmentation/hollow convolution/cascading feature fusion
维普
万方数据