结合生成对抗网络与混合注意力机制的街景图像语义分割

Semantic Segmentation of Street View Images Based on Generative Adversarial Networks and Mixed Attention Mechanisms

街景图像语义分割是自动驾驶领域的主要研究任务之一,对于路径规划和行人安全保障具有重要意义。目前,街景图像语义分割主要存在小目标物体分割不精确、模型容易出现过拟合的问题。为此,提出一种结合生成对抗网络与混合注意力机制的街景图像语义分割模型。具体而言,提出一种多尺度混合注意力模块,用于增强上下文语义信息、提高特征表征能力和对多尺度目标的适应性。同时,为了降低过拟合,引入BN层,结合DCGAN网络构建生成对抗网络分割模型,通过判别损失和分割损失共同约束训练,以增强模型稳定性、提高分割精度。实验结果表明,与DeepLabV3+相比,所提模型在Cityscapes数据集上的分割精度提高了2.4个百分点,mIoU值达到73.4%。

Semantic segmentation of street view images is one of the main research tasks in the field of autonomous driving.At present,seman⁃tic segmentation of street view images mainly suffers from inaccurate segmentation of small target objects and overfitting of models.Therefore,a street view image semantic segmentation model combining generative adversarial networks and hybrid attention mechanisms is proposed.Spe⁃cifically,a multi-scale hybrid attention module is proposed to enhance contextual semantic information,improve feature representation abili⁃ty,and adaptability to multi-scale targets.At the same time,in order to reduce overfitting,a BN layer is introduced and combined with a DC⁃GAN network to construct a generative adversarial network segmentation model.The training is constrained by both discriminative loss and seg⁃mentation loss to enhance model stability and improve segmentation accuracy.The experimental results showed that compared with Deep⁃LabV3+,the proposed model improved segmentation accuracy by 2.4 percentage points on the Cityscapes dataset,with a mIoU value of 73.4%.