Image semantic segmentation method based on improved ERFNet model

Abstract

In order to solve the problems in the existing image semantic segmentation methods, such as the poor segmentation accuracy of small target object and the difficulty in segmentation of small target area, an image semantic segmentation method based on improved ERFNet model is proposed. Firstly, combining the asymmetric residual module and the weak bottleneck module, the ERFNet network model is improved to improve the running speed and reduce the loss of precision. Then, global pooling is used to fuse the feature channels after pyramid pooling to preserve more important feature information. Finally, the network model is implemented based on PyTorch deep learning framework, and the proposed method is demonstrated by experiments, in which the model retraining method is adopted to learn and train it. The experimental results show that the proposed method improves the segmentation ability of small-scale objects and reduces the possibility of misclassification. The average pixel accuracy (MPA) and average intersection merge ratio (MIOU) of the proposed method are higher than those of other contrast methods.