基于沙漏结构的实时语义分割算法

Real-Time Semantic Segmentation Algorithm Based on Sandglass Block

针对实时语义分割网络存在的误分割和目标分割不完整现象,构建了基于沙漏结构的实时语义分割算法(DSGNet)。首先,利用带有空洞卷积的沙漏结构构造骨干网络;其次,使用深度可分离卷积改进空洞空间金字塔模块,并将各分支输出的特征图进行有效堆叠,优化分割效果;最后,设计特征融合模块,高效融合不同维度特征。算法在CamVid数据集进行验证,平均交并比为57.33%,FPS达到了147.3。

Aiming at the mis-segmentation and incomplete target segmentation in the real-time semantic segmentation network,a real-time semantic segmentation algorithm(DSGNet)based on the sandglass block is proposed.First,use the sandglass block with dilated convolution to construct the backbone network.Second,use the depthwise separable convolution to improve the performance of Atrous Spatial Pyramid Pooling module,and stack the feature maps output by each branch effectively to optimize the segmentation effect.Finally,design feature fusion Module to integrate features of different dimensions efficiently.The algorithm was verified on the CamVid dataset,and the mean Intersection over Union was 57.33%,and the FPS reached 147.3.