摘要
图像语义分割是现代自动驾驶系统的一个必要部分,因为对汽车周围场景的准确理解是导航和动作规划的关键。为提高自动驾驶场景的图像语义分割准确率,且考虑到当下流行的基于卷积神经网络的语义分割模型(DeepLab v3+)无法有效地利用注意力信息,导致分割边界粗糙等问题,提出一种融合底层像素信息与通道、空间信息的语义分割神经网络。在卷积神经网络中插入注意力模块,提取出图像语义级别的信息,通过学习图像的位置信息和通道信息得到更加丰富的特征;从卷积神经网络输出的各类别得分值计算出单点势能,且从初步分割图和原图得到成对势能,以便全连接条件随机场对图像的全部像素进行建模,并且优化图像的局部细节;全连接条件随机场通过迭代得到语义分割的最终结果。在CityScapes数据集上进行了测试,与DeepLab v3+相比较,改进后的模型分别提高了均交并比和均像素精度等关键指标1.07%和3.34%。它能够更加精细地分割目标,较好地解决分割边界粗糙,有效地抑制边界区域分割的过度平滑和不合理孤岛等问题。
Semantic image segmentation is an essential part of modern autonomous driving systems because accurate understanding of the scene around the car is the key to navigation and motion planning.The existing advanced convolutional neural network-based semantic segmentation model DeepLab v3+can not use attention information,which leads to rough segmentation boundary.To improve the semantic image segmentation accuracy for autonomous driving scenario,this paper proposed a segmentation model that combined the low pixel information with channel and spatial information.By inserting the attention module in the convolutional neural network,image semantic level information could be extracted,and more abundant features could be obtained through learning the position information and channel information of the image.The unary potential was figured out from the scores of each category output of the convolutional neural network,and the pairwise potential was obtained from the preliminary segmentation and the original input image,so that every pixel of the image could be modeled by fully connected conditional random fields,and the local details of the image could be optimized.The final result of semantic segmentation was obtained from fully connection conditional random fields through iteration.Compared with the existing DeepLab v3+network,the improved model can promote key indicators such as mean intersection over union(mIoU)and mean pixel accuracy(mPA)by 1.07 and 3.34 percentage points respectively.It is able to segment objects more finely,and suppress the excessive smoothness of the boundary region segmentation,unreasonable islands preferably.
作者
王中宇
倪显扬
尚振东
WANG Zhong-yu;NI Xian-yang;SHANG Zhen-dong(School of Instrumentation Science and Opto-Electronics Engineering,Beihang University,Beijing 100191,China;School of Mechatronics Engineering,Henan University of Science and Technology,Luoyang 471023,China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2019年第11期2429-2438,共10页
Optics and Precision Engineering
基金
北京市自然科学基金资助项目(No.3172020)
