Low-speed Metal Droplet Denoising Algorithm Based On Deep Learning

Metal droplets are sub-steady state materials prepared by electrostatic suspension combined with vacuum drop control.To study the dynamic process of droplet nucleation,they are photographed with the help of a camera.Due to equipment,depth of field and environment,most of the images collected are defaced by different noises and low resolution.Image denoising is the removal of unnecessary signals in the image,it is the basis and premise of image processing,and is an important link in the image preprocessing stage.Since there is no efficient denoising network model for the real noise of this scene,this paper explores the image denoising method based on Convolutional Neural Network.Specific work is as follows:Traditional image denoising networks lack image edge details due to too deep layers and lack of attention,and the practical application denoising effect is not good.Considering this problem,this paper proposes an improved Denoising algorithm based on Multi-Scale Channel Concern(MCCD).First,a batch normalization layer is introduced in the residual stacking module to improve the stability of network training,thus solving the problem of slow parameter update caused by inconsistent data distribution.Secondly,an attention mechanism is introduced in the cyclic stacked residual block to transmit richer semantic signals and suppress redundant information,thus improving the denoising ability of the network.Compared with the Deepdeblur network,the algorithm in this paper improves the PSNR by an average of 0.32%,and the SSIM improves by 20%.Given the lack of information interaction in the MCCD network,the denoising effect of metal droplets is not obvious.This paper proposes a denoising algorithm based on multi-scale channel spatial attention and joint loss(MCSD).First,the spatial channel attention mechanism is used to collect the reference information of the deep multi-scale network to achieve feature reusability.Secondly,by constructing a "coarse-fine-fine" multi-scale network to enhance the fusion of key information,transmit multi-scale features to enhance the sensitivity of contextual information,and introduce a spatial channel attention mechanism to better express the details and differences of the droplet image.Finally,the joint loss function is used to train the network.By comparing the data set of this paper,the experimental effect is better than the above method.Compared to the MCCD network,the MCSD network has an average increase of 2.8% in PSNR and 2.27% in SSIM.Finally,Py QT5 and QT Designer are used to build a denoising software system,and metal droplet images are used to verify the effectiveness of the proposed algorithm.