基于轻量化PASMnet神经网络的立体匹配方法  

Stereo Matching Method Based on Lightweight PASMnet Neural Network

作　　者：徐辛超[1] 于佳琪 XU Xinchao;YU Jiaqi(School of Geomatics,Liaoning Technical University,Fuxin 123000,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000

出　　处：《测绘与空间地理信息》2024年第3期8-11,共4页Geomatics & Spatial Information Technology

基　　金：国家自然科学基金项目(42071447、41601494)资助。

摘　　要：针对传统方法用于立体匹配、三维地形重建效果差,而深度学习的方法模型大、速度慢的问题,提出了一种轻量化PASMnet神经网络的立体匹配方法。首先,对PASMnet进行结构化剪枝,精化参数配置,减少模型参数数量,提升模型运行速度。然后对归一化后的图像进行极线校正,得到成像平面平行的立体影像,进行立体匹配任务。在KITTI2015数据集上进行验证,实验结果表明改进的模型参数量减少为原模型的1/6,运行时间为0.56 s。剪枝量化后的模型运行速度更快,推理时间缩短,并且可以用于月面影像的立体匹配,拓展了模型应用。A lightweight PASMnet neural network for stereo matching is proposed for the problems of poor results of traditional methods for stereo matching and 3D terrain reconstruction,and large models and low efficiencies of deep learning methods.First,structured pruning is performed on PASMnet to refine the parameter configuration,reduce the number of model parameters,and improve the model running speed.Then the normalized image is polar line corrected to obtain a stereo image with parallel imaging planes for the stereo matching task.Validated on the KITTI2015 dataset,the experimental results show that the improved model parameter quantity is reduced to 1/6 of the original model,and the running time is 0.56 s.The pruned and quantized model runs faster,has shorter inference time,and can be used for stereo matching of lunar surface images,which extend the model application.

关 键 词：PASMnet 立体匹配 剪枝 量化 月面影像 

分 类 号：P209[天文地球—测绘科学与技术]