基于Transformer的道路场景分割算法研究

Research on road scene segmentation algorithm based on Transformer

图像语义分割技术作为计算机视觉领域的关键技术之一,可以识别并理解图像中每一个像素的内容,并已应用在自动驾驶、医疗诊断、地理信息系统以及图像搜索等很多场景。相对于深度卷积神经网络,Transformer模型基于纯注意力机制,没有任何卷积层或循环神经网络层。本文在Swin Transformer的基础上进行了改进,提出了一种新的网络结构SwinLab。实验结果表明改进后的SwinLab模型相比于深度卷积神经网络的模型算法以及原Swin Transformer模型的分割精度不相上下,mIoU可达80.1,同时在CityScapes数据集上也进行了对比实验,从而进一步证明了该结构的有效性和泛化性。综上,本文在以Swin Transformer为骨干网络的基础上做了相关工作,从而使模型结构更简单,训练和推理速度更快,且准确率也相当可观。

As one of the key technologies in the field of computer vision,images semantic segmentation technology can identify and understand the content of each pixel in an image,and is used in many scenarios such as autonomous driving,medical diagnosis,geographic information systems,and images search.Compared with deep convolutional neural networks,the Transformer model is based on a pure attention mechanism without any convolutional or recurrent neural network layers.Improvements have been implemented on the basis of Swin Transformer,and a new network structure SwinLab is proposed in this paper.The experimental results show that the segmentation accuracy of the improved SwinLab model is comparable to that of the deep convolutional neural network model algorithm and the original Swin Transformer model,and the mIoU can reach 80.1.At the same time,a comparative experiment is also carried out on the CityScapes dataset,so the effectiveness and generalizability of this structure is furtherly demonstrated.In summary,this paper has performed related work on the basis of Swin Transformer as the backbone network,so that the model structure is simpler,the training and inference speed is faster,and the accuracy rate is also considerable.