一种基于对偶学习的场景分割模型

Scene Segmentation Model Based on Dual Learning

城市场景分割等复杂任务存在特征图空间信息利用率低下、分割边界不够精准以及网络参数量过大的问题。为解决这些问题,提出了一种基于对偶学习的场景分割模型DualSeg。首先,采用深度可分离卷积使模型参数量显著减少;其次,融合空洞金字塔池化与双重注意力机制模块获取准确的上下文信息;最后,利用对偶学习构建闭环反馈网络,通过对偶关系约束映射空间,同时训练“图像场景分割”和“对偶图像重建”两个任务,辅助场景分割模型的训练,帮助模型更好地感知类别边界、提高识别能力。实验结果表明,在自然场景分割数据集PASCAL VOC中,基于Xception骨架网络的DualSeg模型的mIoU和全局准确率分别达到81.3%和95.1%,在CityScapes数据集上mIoU达到77.4%,并且模型参数量减少18.45%,验证了模型的有效性。后续将探索更有效的注意力机制,进一步提高分割精度。

For complex tasks such as urban scene segmentation,there are problems such as low utilization of feature map space information,inaccurate segmentation boundaries,and excessive network parameters.To solve these problems,DualSeg,a scene segmentation model based on dual learning,is proposed.Firstly,depthwise separable convolution is used to significantly reduce the number of model parameters Secondly,accurate context information is obtained by fusing hollow pyramid pooling and double attention mechanism modules.Finally,dual learning is used to construct a closed-loop feedback network,and the mapping space is constrained by duality,while training the two tasks of“image scene segmentation”and“dual image reconstruction”,it can assist the training of the scene segmentation model,help the model to better perceive the category boundary and improve the recogni-tion ability.Experimental results show that the DualSeg model based on the Xception skeleton network achieves 81.3%mIoU and 95.1%global accuracy on natural scene segmentation dataset PASCAL VOC,respectively,and the mIoU reaches 77.4%on the CityScapes dataset,and the number of model parameters decreases by 18.45%,which verifies the effectiveness of the model.A more effective attention mechanism will be explored in the future to further improve the segmentation accuracy.