Study On Denoising Algorithm Of Grain Depot Monitoring Image Based On Deep Learning

In order to adapt to the development of the times and to better manage food security issues,the construction of "intelligent grain depot" is booming.Intelligent security,target identification and tracking equipment in the grain depot security has also been widely used.The intelligent security video monitoring system can monitor the working conditions of some important places in the grain depot in real time,such as the main import and export grain storage channels,storage areas,work points,equipment depots,drug depots and so on.At the same time,it can also warn against abnormal behaviors such as personnel gathering,crossing the border,regional invasion and violation of rules and regulations by operating personnel,which not only reduces on-site inspection,but also reduces the cost of manual management of grain depots,more conducive to the process of grain storage management and data collection.However,due to the limitations of the hardware conditions and the harsh environment of the various image acquisition equipment in the grain dearies,noise is inevitably generated in the process of image acquisition,which will have a negative impact on the accuracy of the subsequent image data analysis and processing results.In addition,the grain depot image as the data set contains less target feature types and the feature data set is not rich enough.Therefore,on the basis of in-depth analysis of deep learning image denoising algorithm,this paper conducts a comprehensive study on how to effectively filter out the noise of grain depot images and minimize the loss of edges and details of grain depot images,so as to obtain higher quality and clearer grain depot images.The specific research contents are as follows:(1)An improved depth residuals denoising algorithm based on multi-scale and atruos convolution(MAC-Dn CNN)is proposed,which is based on Dn CNN image denoising network.In the MAC-DNCNN algorithm,the multi-scale feature extraction network layer is used to extract enough shallow features from the input noisy image,so as to make up for the shortcomings of less available target information and less rich feature domain in the grain depot image,and increase the adaptability of the network to scale.At the same time,atrous convolution and jump connection technology are used to design Resnet Unit residual unit block to realize residual learning,which can better learn noisy residual images,so as to remove image noise and achieve the purpose of image clearness.(2)An improved image denoising algorithm(SWT-GAN)combining wavelet transform and generative confrontation network is proposed.Considering that wavelet transform processing can effectively separate the basic information of low-frequency image from high-frequency noise information in the image,and compared with all other models,GAN can produce clearer and more real samples.In view of the problem that the image denoising algorithm MAC-DNCNN proposed in(1)has a higher value of objective index after denoising,but the texture details are still not clear enough and need to be improved,a generator and discriminant structure of SWT-GAN network is designed.The main idea is to input the three high-frequency coefficient maps(main noise areas)obtained from the noisy image wavelet decomposition into the SWT-GAN model for processing,and then combine them with the wavelet decomposition obtained by the wavelet decomposition of the SWT-GAN image denoising network The low-frequency coefficient map(the area where the basic information of the original image exists)is reconstructed together to obtain a high-quality,clear image after denoising.(3)The improved SWT-GAN image denoising algorithm proposed in(2)is to make the denoised image contain more original image details and edge information,in this paper,we introduce a new loss function(Perceptual loss function)into SWT-GAN network model.The perceptual loss can add some feature information extracted from the convolutional neural network into the training process of the network,and constrain the training process as part of the objective function to guide the learning direction of the network.In the process of training,the differences between the generated image and the target image on the feature level are reduced,so that the generated image can have a higher consistency with the target image,and then the reconstructed image has a higher definition.According to the characteristics of grain depot images,the improved algorithm of deep residual denoising network based on multi-scale and atrous convolution(MAC-Dn CNN)and the improved algorithm of image denoising(SWT-GAN)combined with wavelet transform and generating confrontation network proposed in this paper are both can achieve ideal results in denoising processing of grain depot images.