Abstract
In the paper, we propose a generalized deep nonlinear CNN model, named as MDD-Net, for filtering various kinds of ESPI fringe patterns with structure protection and shape preservation. We firstly design the multi dilated dense module (MDD-Module) via merging five 1-, 2-, 5- dilated convolutions with abundant skip connections. Then, we construct the MDD-Net via combing three MDD-Modules and five common convolutions in a staggered series manner with skip connection. In this way, the MDD-Net enhances the reuse of original features, improves the flow of deep features, promotes the fusion of different deep features, finally obtains the rich deep features in the filtering process. We also propose a shape-structure-consideration loss function by combining the MAE, MSE and SSIM loss functions, and construct an available pertinent dataset for ESPI fringe filtering. With the proposed dataset and the proposed loss function, we train the MDD-Net successfully. With the trained network, we directly get the results in batch without any parameter fine-turning and any pre-process or post-process. We test our method on many simulated and experimental ESPI fringes with different densities, and compare it with the existing SOOPDE, WFF, BL-Hilbert-L2, and FFD-Net methods. The performance is also evaluated quantitatively and qualitatively in terms of speckle reduction, structure protection, shape preservation, parameter finetuning and running time. Results demonstrate that our method can simultaneously reduce speckles, protect structures and preserve shapes in the fringe filtering and get the better results than the compared methods in all cases. Moreover, it has full advantages on excellent generalization and batch performance, and can be used to process a great number of ESPI fringe patterns rapidly. In fact, it has been successfully applied to the dynamic measurement in the paper.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call