Image Super-Resolution Reconstruction Based On Deep Learning And Its Application In The Detection Of PCB Welding Quality

In practical applications,image is one of the most important information carrier.With the increasing demand for image quality,obtaining high resolution images with higher quality has gradually attracted more and more attention from the public.Improving hardware devices is the most direct way to increase image resolution,but it will cause problems such as increased costs,wasted resources,and even if the improvement of equipment cannot meet the demand in some occasions.Therefore,the use of image super-resolution reconstruction technology to improve the image resolution has become an effective choice.It is low-cost,simple and convenient,and it has good results and a wide range of applications.This thesis mainly studies on image super-resolution reconstruction based on deep learning,researches two improved image super-resolution reconstruction methods,and applies them to the detection of PCB welding quality.Firstly,because of the slow training speed and insufficient information in existing image super-resolution reconstruction methods,this thesis studies on a super-resolution reconstruction method based on multi-band convolutional neural network image.This method adopts the concept of residual learning,uses low-resolution images and the information in the low-resolution feature space to generate low,medium,and high frequency band information through different routes,and uses the same reconstruction layer to obtain the high resolution image.This method can speed up training process,make full use of image information and reconstruct higher resolution images.Secondly,although the first image super resolution method has a fast training speed and a good reconstruction performance because of its simple structure,the image details cannot be fully reconstructed.To solve these problems,an image super-resolution reconstruction method based on a multi-band of memory transfer deep convolutional neural network has been studied in this thesis.The concept of memory transfer is derived from the residual learning.The network forms a multi-band learning structure with local and global,which leads to a new type of deep convolutional neural network.This method makes full use of image information and has good reconstruction performance and high efficiency.Thirdly,this thesis also analyzes the parameters of the network separately.Furthermore,this thesis applies the above two image super-resolution reconstruction methods to the detection of PCB welding quality,which obviously improves the PCB image resolution and is conducive to the detection technology of PCB welding quality.Experiments reveal that the two image super resolution reconstruction methods studied in the thesis can reconstruct high resolution images with better quality compared with other representative methods,and improve the objective evaluation indicators and visual results too.Moreover,the two methods also improve the image quality in the application of PCB images significantly,which can be beneficial for the next operation of image processing in the detection system.