基于奇偶交叉卷积的轻量级图像语义分割网络

传统轻量级图像语义分割网络中的跨步下采样卷积导致感受野呈现刺状分布，使得像素点利用率出现系统偏差，影响分割精度的提升。为此，针对传统的轻量级图像语义分割网络设计一种奇偶交叉卷积的下采样模块，在跨步奇数卷积模块前增加单步偶数卷积模块，在一定程度上缓解了刺状分布带来的不良影响，消除了分割网络中不同空间位置上像素利用率的偏差，最终提高了模型对像素点的分割精度。通过7种不同轻量级图像语义分割网络的对比可以看出，所提模型可以明显消除刺状分布，使分割网络的精度进一步提高，同时也证明了该模型适用于不同的轻量级网络，具有普适性。

Lightweight Image Semantic Segmentation Network Based on Parity Cross Convolution

The multi-step down-sampling convolution in traditional lightweight image semantic segmentation networks easily causes a thorn-like distribution in receptive fields. This will introduce systematic deviations in pixel utilization and finally affect the improvement of segmentation accuracy. Regarding this problem, a down-sampling module of parity cross convolution is designed for traditional lightweight image semantic segmentation networks. In the scheme, an even-convolution module is added before the strided odd-numbered convolution module, in order to alleviate the negative effects caused by spur distribution. Accordingly, the deviation of pixel utilization at different spatial positions in the segmentation network can be eliminated effectively, and the improvement of pixel segmentation accuracy can be finally achieved. Experimental results demonstrate that through the comparison of seven different lightweight image semantic segmentation networks, the proposed model can obviously eliminate the thorn-like distribution and improve the accuracy of the segmentation network. Also, the proposed model performs excellent adaptability to different lightweight networks.