基于改进BiSeNet的室内场景语义分割方法

A Method of Semantic Segmentation for Indoor Scenes Based on Improved BiSeNet

室内场景的语义分割一直是深度学习语义分割领域的一个重要方向。室内语义分割主要存在的问题有语义类别多、很多物体类会有相互遮挡、某些类之间相似性较高等。针对这些问题,提出了一种用于室内场景语义分割的方法。该方法在BiSeNet(bilateral segmentation network)的网络结构基础上,引入了一个空洞金字塔池化层和多尺度特征融合模块,将上下文路径中的浅层细节特征与通过空洞金字塔池化得到的深层抽象特征进行融合,得到增强的内容特征,提高模型对室内场景语义分割的表现。该方法在ADE20K中关于室内场景的数据集上的MIoU表现,比SegNet高出23.5%,比改进前高出3.5%。

Semantic segmentation of indoor scenes has always been an important direction in the field of deep learning semantic segmentation.The main problems of indoor scene segmentation are many semantic categories,many object classes will block each other,and some classes have high similarity.Aiming at these problems,Proposed a method for semantic segmentation of indoor scenes which is based on the BiSeNet(bilateral segmentation network),this method introduces a hollow pyramid pooling layer and a multi-scale feature fusion module.The features are fused to obtain enhanced content features,which improves the model s performance for semantic segmentation of indoor scenes.The MIoU performance of this method on the indoor scene dataset in ADE20K increased by 23.5%compared toSegNet and 3.5%compared to before model.