Panoramic image semantic segmentation using channel attention-based HarDNet and distorted boundary learning

In this paper, we propose a semantic segmentation framework for panoramic images. First, in order to solve the problem of large panoramic image size, we use HarDNet in the backbone. By applying HarDNet, while improving segmentation accuracy, it also improves the training efficiency. Secondly, we embed the efficient channel attention mechanism to solve the problem that different channels of the feature map occupy different importance in the convolution pooling process. Finally, boundary loss function is introduced in the training process to make the method pay more attention to the edge features of the object. This helps to further improve the effect and accuracy of panoramic semantic segmentation, so that the proposed method can better capture the contour information of the object. To evaluate the performance of the improved algorithm, we tested on the public dataset of Stanford 2D-3D-Semantics. The experimental results show that the mAcc and mIoU of the improved algorithm reach 54.1% and 66.7%, outperforms other algorithms.