Research On Motion Deblurring Of Industrial Character Image Based On Generative Adversarial Network

Image is the basis of human vision and contains rich semantic information.In industrial applications,obtaining clear images is the premise of target recognition and defect detection.It will be more challenging to motion deblurring from multi-textured images such as character images.Thanks to the excellent extraction and flexible expression of image feature information by the convolutional neural network,the studies of image motion deblurring based on deep learning have progressed.Therefore,to meet the practical application requirements applied in the industrial field,maximize the restoration of texture details of motion-blurred character images,and ensure the promotion of the normalization detection process,this paper will study the motion deblurring of industrial character images based on generative adversarial network.The main contributions are summarized as follows:1.A lightweight motion deblurring algorithm based on depthwise separable convolution is proposed.Firstly,aiming at the problem that the parameters of the convolutional neural network model are too large to meet the industrial application,a lighter feature extraction method is proposed in this paper.On the premise of ensuring the accuracy of feature extraction,this method is based on depthwise separable convolution,pays attention to executing the correlation facing the interior of the kernel,and uses fewer parameters to extract image features.Secondly,in building the model,a bi-linear interpolation and convolution up-sampling method(BC)is used to avoid the overlapping phenomenon caused by the deconvolution operation.The experimental results show that the SandglassDepthwise Feature Extraction Block(SDB)has fewer parameters and strong feature information extraction ability than other lightweight modules.The performance of the light motion deblurring method based on depthwise separable convolution has been further improved,which meets the standard of industrial application.2.The research target is focused on multi-textured images in this stage.According to the characteristics of such images,a Texture Details Oriented Generative Adversarial Network(TDGAN)is proposed.On the one hand,fine-grained convolution and multi-scale channel fusion strategies are used to strengthen feature information extraction,and channel shuffling strategy is used to ensure the extraction balance of information.On the other hand,the sparse perceptual loss function is optimized.The optimized loss function is integrated into the model to make the generated multi-textured image closer to the original in edge detail and brightness.Many experiments on public dataset Go Pro and characters dataset Ass IMG show that the algorithm is superior to other algorithms in texture details preservation and edge clarity of textures,which results in image quality evaluation indexes PSNR and SSIM.3.The proposed method,TDGAN,is used in the industrial scene to verify the above method because of the multi-textured industrial packaging character image’s characteristics.One Normalized detection scheme of packaging character images based on template matching and the adaptive region location method based on feature region is proposed in the paper.