Global Feature-Guided Real-Time Semantic Segmentation Algorithm

With the development of convolutional neural networks, a large number of high-precision semantic segmentation algorithms have been proposed. However, these algorithms are too complex to infer images in time, and difficult to apply in industrial scenarios. To address this problem, in this paper we therefore propose a lightweight real-time semantic segmentation network which is guided by global features. We first design an inverted residual module to extract the features of the image. Then, the phased feature aggregation module based on the global pooling layer is employed to fuse the features to transmit information from the low level to the high level. Finally, in the feature recovery stage, the image boundary extracted by the Laplacian of Gaussian operator provides boundary information for the deep features to solve the problem of the loss of detailed information, such as boundaries in the continuous down-sampling process of image features. The Experimental results on the CamVid dataset show that the proposed method has a good performance in the inference speed and segmentation accuracy. The inference speed of processing 960×720 images can reach 210FPS, and the segmentation accuracy can reach a mIoU of 68.6%.