基于知识蒸馏和超分辨率的车道检测算法仿真  

Simulation of lane detection algorithm based on knowledge distillation and super-resolution

作　　者：张琰[1,2] 舒畅 王晶 ZHANG Yan;SHU Chang;WANG Jing(Information Engineering Institute,Wuchang Institute of Technology,Wuhan Hubei 430065,China;Universitéde Technologie de Compiègne,dans le department de l′Oise en région Picardie,Compiègne 60200,France)

机构地区：[1]武昌工学院信息工程学院,湖北武汉430065 [2]法国贡比涅技术大学(UTC),皮卡第大区瓦兹省贡比涅60200

出　　处：《计算机仿真》2023年第1期178-181,194,共5页Computer Simulation

基　　金：湖北省教育厅科学技术研究计划指导性项目(B2020266);湖北省高等学校优秀中青年科技创新团队计划项目(T2021042)。

摘　　要：为改善自动驾驶领域中车道线检测的鲁棒性与实时性,提出了基于知识蒸馏和超分辨率的快速车道线检测算法。由于复杂路况与采集图像的分辨率因素影响,会导致对空间和边缘信息的缺失,而单纯利用神经网络来提高图像的特征提取与分割精度,又会导致网络结构过于繁杂。因此结合知识蒸馏思想对神经网络采取轻量化设计,并对路况图像采取超分辨率重构。算法首先采用分组卷积,把Teacher网络的训练成果转换成不同分辨率的投影关系,完成缺失信息的合理补充。针对Teacher与Student网络相似性,为避免训练过程中的相互干扰,引入残差结构进行特征重建。根据编解码设计分割网络,通过特征提取与交叉熵计算,确定车道线的分布情况。基于Cityscape数据集对算法的性能进行比较分析,结果表明,所提算法能够有效提高车道线检测的实时性,同时具有良好的鲁棒性和准确率。In order to improve the robustness and real-time of lane line detection in the field of automatic driving, a fast lane line detection algorithm based on knowledge distillation and super-resolution is proposed. The influence of complex road conditions and the resolution of collected images can lead to the lack of spatial and edge information, and simply using neural network to improve the accuracy of image feature extraction and segmentation can lead to too complicated network structure. Therefore, combined with the idea of knowledge distillation, the lightweight design of neural network was adopted, and the super-resolution reconstruction of road condition image was adopted. Firstly, we used grouping convolution to convert the training results of teacher network into projection relations with different resolutions to complete the reasonable supplement of missing information. Aiming at the similarity between teacher and student networks, in order to avoid mutual interference in the training process, the residual structure was introduced for feature reconstruction. Then, the segmentation network was designed according to the codec, and the distribution of lane lines was determined through feature extraction and cross entropy calculation. Finally, the performance of the algorithm was compared and analyzed based on cityscape data set. The results show that the proposed algorithm can effectively improve the real-time performance of lane detection, and has good robustness and accuracy.

关 键 词：知识蒸馏 超分辨率 神经网络 损失函数 车道线检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]