DFNet: Towards Real-Time Semantic Segmentation with Depthwise Separable Convolution

Semantic segmentation refers to assigning a category label (such as pedestrian, car, building, road, sky) to each pixel in the image to help the computer understand the content of the image and conduct subsequent processing, such as the automatic driving car to identify roads, vehicles, pedestrians from the acquired images to make correct driving decisions, or help doctors diagnose diseases in medical image analysis. It also can provide semantic information for robot navigation. Running semantic segmentation algorithms in edge devices requires powerful hardware performance, most of the current deep neural networks is large, which seriously limits their application and deployment in practical scenarios. For this purpose, we propose a new convolutional network architecture named DFNet, which employ dilated convolution and depthwise separable convolution into segmentation networks, aiming to find the best balance between accuracy and speed. Tests on cityscapes dataset show that our model is almost as accurate as UNet, but model parameters are reduced by 4 times and the segmentation speed is increased 75%.