| Image semantic segmentation technology,as one of the means for autonomous vehicles to perceive the environment,has made great progress in recent years with the development of deep learning.It is essentially a pixel-level intensive prediction task,which often consumes a lot of computing resources and is difficult to achieve real-time application.In this paper,the following three aspects are accomplished to address these problems.Firstly,two lightweight fast residual basic modules are designed and their advantages are analyzed in terms of the number of parameters,the perceptual field and the relative proportion of convolution kernels of different sizes.Then based on the two blocks,a lightweight feature extractor with fewer number of parameters and a fast feature extractor with faster inference speed are constructed.The shallow part of both models uses a dual-path fast downsampling module,the main part uses fast residual basic modules with three-stage stacked blocks,and the output part uses a fully connected layer.Finally,two efficient real-time semantic segmentation algorithms with asymmetric encoder-classifier of decoding structure are proposed.In the algorithms,a lightweight mutual attention feature fusion algorithm that can combine both high and low level channel feature information and overcome the computational difficulties caused by different resolutions is designed to effectively fuse high and low level feature information? a lightweight atrous spatial pyramid pooling module and a pyramid pooling module with channel pooling information are designed to faster and better capture multi-scale contextual information.The experiments show that two feature extractors have about 1～2% accuracy advantage over other models on the CIFAR10,and have speed advantages of 30～50 frames and 60～90 frames,respectively.The two segmentation models are able to achieve 68.4%MIo U,67.5 FPS and 67.1% MIo U,105.0 FPS performance on the Cityscapes test set,respectively. |
PDF Full Text Request