|
In recent years, the convolutional neural network has achieved remarkable success in semantic segmentation of traffic scene understanding. At present, the main problems in the field of semantic segmentation are as follows: 1) The repeated pooling and downsampling operations reduce resolution of traffic images in the convolutional networks, which leads to lose abundant spatial information and poor segmentation performance. 2) Traffic images contain many objects of different scales. How to accurately recognize and segment these multi-scale objects is another key problem in semantic segmentation. To handle these problems, this paper propose an image semantic segmentation method based on the Residual Dilated Attention. This method uses spatial CNN to extract high-level semantic information, and then uses the proposed model to capture low-level semantic information, and follows the designed sampling rules to set appropriate and effective sampling rates, and effectively aggregates multi-scale context information while maintaining high resolution of feature maps. Finally, this paper also designs a fusion module to effectively fuse the results generated by the spatial CNN and the Residual Dilated Attention. The method in this paper conducts a series of simulation experiments on CULane and CamVid traffic datasets, and achieves competitive results, proving the effectiveness of the proposed method. |
|
Keywords:Computer vision; Semantic segmentation; Attention mechanism; Dilated convolutions; Multi-scale context information. |
|