面向语义分割模型的外接多尺度投票网络

针对语义分割领域中多尺度共享网络训练复杂度高，以及网络在小目标、长条状目标、目标边缘处拟合效果不佳的问题，提出一种新型外接多尺度投票网络。通过投票网络融合各尺度分割结果，降低网络训练复杂度，并将共享网络中的分割网络与各尺度注意力头剥离开，仅训练各尺度注意力头，以便于网络收敛。在投票网络的结构设计中，使用多类别投票方法扩大投票空间，通过融入混合池化模块聚合近程与远程权值，扩大网络感受野，缓解权值图中长条状目标拟合间断与缺失的问题。在此基础上引入类内、类间投票注意力模块获取权值及类间关系，并采用不规则卷积，改善投票权值图的边缘拟合效果。在Cityscapes数据集上的实验结果表明，相比FCN、PSPNet、DeepLabv3+分割网络，该网络的平均交并比分别提升了0.92、0.88、0.80个百分点，与共享网络相比，其训练复杂度更低，精度更高。

External Multi-Scale Voting Network for Semantic Segmentation Model

To address the high training complexity of multi-scale shared networks in semantic segmentation and the unsatisfactory fitting effect of multi-scale attention maps for small targets, long targets, and target edges, a new external multi-scale voting network is proposed.The segmentation results of each scale are combined via the voting network to reduce the training complexity.Moreover, the segmentation network in the shared network is separated from the attention heads of each scale, and only the attention heads of each scale are trained to facilitate the convergence of the network.The multi-category voting method is used to expand the voting space.By integrating the hybrid pooling module to aggregate short and long-range weights, the receptive field of the network is expanded, and the discontinuity and missed fitting of long bar-shaped targets in the weight map are alleviated.Subsequently, intra-class and inter-class voting attention modules are introduced to obtain the weights and inter-class relations, and irregular convolution is used to improve the edge-fitting effect of the voting weight map.Experimental results on the Cityscapes dataset show that the mean Intersection over Union (mIoU) of the proposed network is higher than those of FCN, PSPNet, and DeepLabv3+ segmentation networks by 0.92, 0.88, and 0.80 percentage points, respectively.Compared with a shared network, the proposed multiscale voting network offers higher accuracy and requires less complex training.