CFF-SEgFormer: Lightweight Network Modeling Based on Segformer

In automatic driving, previous road scene semantic segmentation algorithms tend to increase the number of parameters and computation to improve accuracy. This results in slow semantic segmentations of road images not meeting the real-time requirements of automatic driving. In this study, based on the transformer hybrid transformer encoder, a lightweight decoder with high accuracy and small computational volume with cross-feature fusion was developed to replace the traditional multilayer perception (MLP) decoder. The results of the experiment showed that the model improved the mIoU of the traditional MLP decoder from 73.31 to 75.43%, the Dice coefficient from 84.0 to 85.4
^%
, and reduced the computation amount from 71.48 G to 39.27G on the Cityscapes dataset without increasing the number of parameters.