Center for Intelligent and Interactive Robotics, Korea Institute of Science and Technology, Seoul 02792, Republic of Korea.
Department of Mechanical Convergence Engineering, Hanyang University, Seoul 04763, Republic of Korea.
Sensors (Basel). 2022 Dec 28;23(1):333. doi: 10.3390/s23010333.
We present a new deep learning framework for removing honeycomb artifacts yielded by optical path blocking of cladding layers in fiber bundle imaging. The proposed framework, HAR-CNN, provides an end-to-end mapping from a raw fiber bundle image to an artifact-free image via a convolution neural network (CNN). The synthesis of honeycomb patterns on ordinary images allows conveniently learning and validating the network without the enormous ground truth collection by extra hardware setups. As a result, HAR-CNN shows significant performance improvement in honeycomb pattern removal and also detailed preservation for the 1961 USAF chart sample, compared with other conventional methods. Finally, HAR-CNN is GPU-accelerated for real-time processing and enhanced image mosaicking performance.
我们提出了一种新的深度学习框架，用于去除光纤束成像中包层光学路径阻塞产生的蜂窝状伪影。所提出的框架 HAR-CNN 通过卷积神经网络 (CNN) 提供了从原始光纤束图像到无伪影图像的端到端映射。在普通图像上合成蜂窝图案可以方便地通过额外的硬件设置进行学习和验证网络，而无需进行巨大的真实数据收集。结果表明，与其他传统方法相比，HAR-CNN 在去除蜂窝图案方面表现出显著的性能提升，并且对 1961 年美国空军图表样本也有很好的细节保留。最后，HAR-CNN 进行了 GPU 加速处理，以实现实时处理和增强的图像拼接性能。
