深度学习密集匹配中样本精度与匹配精度的关系研究

Relationship Between Sample Accuracy and Dense Matching Accuracy in Deep Learning

随着计算机硬件的发展,深度学习密集匹配方法在近景数据集上取得了惊人的效果,但密集匹配监督方法所需的样本标注困难(尤其是航空影像),且数据区域的样本可能存在误差,而这些误差对密集匹配精度的影响未知。针对该问题,通过模拟样本系统误差、随机误差和粗大误差的方式分析了各项误差与密集匹配精度的关系。结果表明:①迁移学习法可有效提高密集匹配精度,使误差平均减少15.3%,并能提升网络抗噪能力,加速网络收敛,减少训练时间;②深度学习网络在一定范围内具有容错能力,但误差增幅会随噪声平均偏移量的增加而逐渐变大;③深度学习的抗噪能力主要体现在随机误差方面,系统性整体偏差对精度影响更大,尤其是基于视差大小的百分比误差,将大幅降低匹配精度。通过进一步分析Vaihingen与WHU数据集增加系统误差后的表现发现,对于环境复杂且数据样本较少的情况,系统误差可能导致网络直接训练不收敛,此时可采用迁移训练的方法改进。

Dense matching algorithms based on deep learning have achieved significant results with close range datasets.However,the data required for dense matching supervision are difficult to acquire.In particular,aerial images may contain errors,which exert an unknown influence on dense matching accuracy.To address this problem,we analyzed the relationship between errors and the dense matching accuracy by simulating sample errors,including system,random and gross errors.The results demonstrate that①the transfer learning method can effectively improve the dense matching accuracy,reduce errors by 15.3%on average,increase the network antinoise capacity,accelerate network convergence,and reduce the training time.②The deep learning network exhibits a certain faulttolerant ability.However,the error margin gradually increases with an increase in the mean noise offset.③The deep learning antinoise capacity is primarily manifested to resist random errors,and the systematic deviation significantly affects the matching accuracy.In particular,the percentage errors based on the disparity value substantially reduce the matching accuracy.Further analysis of increased system errors in the Vaihingen and WHU datasets demonstrates that in complex environments,discrete disparity distribution and with small sample sizes,system errors may cause direct training network non-convergence,which can be addressed through transfer training.