Research On ESPI Image Denoising Based On Deep Learning

In the process of information transmission,image is one of the most important media carriers.The quality of image determines the efficiency and accuracy of information transmission to a large extent,but the existence of image noise has always been a key factor affecting image quality,so the image denoising is the key step of image processing.The electronic speckle pattern interferometry(ESPI)is one of the most effective ways for rough surface deformation measurement,vibration survey and internal damage detection due to its simple operation and convenient real-time monitoring.At the same time,it has the advantage that the detection result is less affected by external interference.The fringe structure of the ESPI image contains the key information about object deformation and displacement.However.a lot of noise will inevitably be generated in the process of obtaining the image,which will seriously interfere with the acquisition of effective information,therefore ESPI image denoising has high research valueTraditional image denoising methods have many characteristics,for example,it is diverse,simple in principle and convenient in process,but the denoising result is not ideal,especially the noise in ESPI image is high and its variety is complex.The denoised image still has many problems.On the one hand,the noise can't be completely removed and there is a certain level of noise residue,which will cause serious interference on the key information.On the other hand,the traditional denoising method will cause significant damage to the image and cannot ensure the integrity of the image information.And the fringe information will be lost to a certain extent,which will be not conducive to further information extraction research.Nowadays,deep learning research has made great progress.Research shows that the neural network denoising model has achieved better denoising effect than the traditional filtering methods.This thesis makes use of the convolutional neural network denoising model,determines the optimal simulation parameters of the simulated ESPI image in network model training,combines the BM3D(Block-Matching and 3-D filtering)-adaptive TV(Total Variation)algorithm with the convolutional neural networks(CNN)model,and proposes an improved CNN denoising model based on BM3D-adaptive TV filtering.It can be seen from the experiments that this model can effectively reduce the noise residue in the traditional methods and can achieve better denoising result for image with high degree of noise.At the same time,it can effectively guarantee the integrity of the image information,avoid the loss of the effective information in the original image,and protect the fringe edge information well,which provides a better working basis for subsequent studies like phase information extraction.The main innovations of this thesis are1.The deep CNN denoising model is used for the ESPI image denoising.The denoising model is trained by simulated images,and the optimal ESPI image simulation method and the optimized network parameters are obtained through comparative experiments.The experimental results show that this method has better denoising effect than other methods,in which the image fringe and edge information are well protected to avoid the loss of effective information.But there is still a small amount of noise residue in the denoised image.2.BM3D-adaptive TV filtering and neural networks method are combined to propose an improved two-channel convolutional neural networks denoising method based on BM3D-adaptive TV filtering.Simulated and real images are fed into two different channels respectively to train the model for real ESPI image denoising.Experimental results show that the proposed method can improve the denoising effect under the premise of ensuring the integrity of image information and avoiding the loss of stripe detail information,which can avoid the influence of the residual noise on the subsequent research like phase extraction.