Semantic Segmentation of Road Traffic Sign based on Improved Deeplabv3+

Road traffic sign is an important facility to manage traffic and indicate the direction of traveling to ensure smooth road and driving safety. However, sign is usually a small target, which is prone to the problem of missed detection and false detection. In addition, the deeplabv3+ model is a representative semantic segmentation network. It employs dilated convolution in the ASPP module, which is prone to losing small target information while expanding the sensory field. Aiming at the above problem, an improved deepswin small target semantic segmentation based on deeplabv3+ is proposed. First of all, the ASPP module is replaced by swin-transformer block to enhance the signage feature extraction capability. Then, a channel and spatial fusion attention(CSFA) mechanism based on CBAM attention mechanism is utilized to enhance the extraction capability of channel and spatial features. The experimental results show that compared with the original network, the MIOU and mPA of the network proposed in this paper are improved by 4.0% and 4.9%, respectively.