Research On Particle Image Velocimetry Algorithms Based On Deep Optical Flow Learning

Particle image velocimetry(PIV)technique is able to extract the velocity field information of fluid from the successive images,which can help researchers get a deeper insight of complicated flow phenomena.Therefore,PIV technique is widely used in the research of experimental fluid mechanics and many other fields.However,the existing PIV algorithms(cross-correlation method and optical flow approaches)still have shortcomings,which are difficult to meet the requirements of complex flow motion estimation.Specifically,cross-correlation approach has the disadvantages in prediction accuracy and spatial resolution,and the optical flow approach is sensitive to the varying illumination and quite time-consuming.Recently,deep optical flow learning methods have made great progress in the field of motion estimation.Although there are few studies based on deep optical flow learning methods for flow field measurement,the current research work has proved that deep optical flow learning is a promising alternative to PIV algorithms.Further analysis shows that the existing PIV algorithms based on deep learning have problems such as redundant model parameters and insufficient precision.In addition,the flow scene has complex and variable characteristics.At present,researchers have not explored the application of deep learning for PIV estimation in more complex flow scenes.Therefore,the research on particle image velocimetry algorithm based on deep learning has important practical significance and application value.In this paper,we conducted in-depth research on the various problems mentioned above,and proposed the PIV approaches based on deep optical flow learning for different task scenarios.The proposed approaches effectively improve the accuracy,spatial resolution,robustness,and computational efficiency of PIV estimation.Overall,the main research work of this paper contains the following four aspects:First,the existing PIV deep learning models have the problems of redundant parameters and insufficient precision.To solve these problems,we propose a lightweight optical flow model called Light PIVNet for PIV estimation.We first adopt a siamese feature encoder with input resolution of 1/4 to efficiently extract features from particle images pair.Then,a 4D correlation volume is constructed to perform visual similarity computation on the extracted image features.Afterwards,the recurrent neural network(GRU)is used to update the flow fields,which can predict the velocity field through multiple recurrent iterations.By this means,it is able to effectively avoid the increase of model parameters.Finally,to further improve the accuracy and generalization ability of the model,we generated a new PIV dataset,which increases the amount and variety of flow fields.The experimental results show that the proposed Light PIVNet effectively improves the estimation accuracy,and has the advantage of lightweighting with fewer model parameters.Second,in the PIV images containing occluded structures,the structures cause great interference to the calculation of the velocity field in the liquid phase region.To solve this issue,we propose a cascaded deep learning framework(termed PIV-UNet+Light PIVNet-2P)to automatically perform the masking of the non-liquid region and estimation of the velocity field in the liquid region.We first use the image segmentation network PIV-UNet to mask the non-liquid area to accurately extract the measured liquid area.The input of the Light PIVNet model is then modified,called Light PIVNet-2P,to automatically calculate the velocity field of the extracted liquid phase region.In addition,the corresponding image segmentation dataset and PIV dataset are constructed for the training of the model.Experimental results show that our proposed approach can accurately extract liquid phase from PIV images containing occluded objects,and achieve high-precision velocity field predition for liquid phase region.Third,uneven laser illumination will cause the changes in the brightness of the PIV image pair,leading to the problem of illumination sensitivity and poor robustness of the algorithm.To solve this issue,a deep optical flow learning model called PIV-Lite Flow Net-D is proposed to resist illumination changes.We first employ an optical flow architecture with a six-level feature pyramid structure for full-resolution flow field estimation.Then the prior assumption knowledge such as brightness gradient consistency and divergence-curl consistency is coupled to the multi-layer training loss function of the feature pyramid to guide the model training.In addition,a light intensity PIV dataset simulating various illumination changes is constructed for model training and testing.Experimental results show that our proposed model is more robust than traditional optical flow algorithms,and can get high-resolution and high-precision velocity fields compared to cross-correlation algorithms.Fourth,the calculation of the velocity field of PIV video(i.e.,multi-frame images)has the problems of poor real-time performance and low calculation efficiency.To address this issue,we put forward a spatio-temporal recurrent estimation model called Deep-TRPIV for multi-frame PIV estimation.We first build a double-frame PIV estimation unit based on Light PIVNet,which uses the convolutional GRU module to iteratively estimate the spatial flow field.Then,a time-recurrent multi-frame PIV estimation model is constructed by adopting feature fusion and flow transfer,and it can efficiently predict the velocity field of N-frame successive particle images.In addition,a multi-frame PIV dataset are generated to train and optimize model parameters.Experimental results demonstrate that our proposed model can effectively predict N-1 high-precision and high-resolution velocity fields when fed N images,and significantly improve the computational efficiency.