| With the rapid development of computer vision and deep learning technology,various online detection and recognition technologies based on video images have been widely used in scientific research,medicine,industry and security detection.However,these are premised on high quality input images.When there are problems such as insufficient illumination,noise interference or insufficient resolution,the practical application effect will be seriously affected.Based on this,this paper carries out research work on low-quality image enhancement and super-resolution restoration algorithm based on convolution neural network.For low-quality images such as low illumination,noise and low-resolution,the convolution neural network is used as the main tool to carry out relevant research work,including:A low-illumination image enhancement algorithm based on U-shaped network is designed and implemented.Under the basic structure of Unet network,jump-link fullconvolution neural network is used to model the non-linear mapping relationship between high-and low-illumination images.By sharing parameters,the parameters are reduced.After convolution layer,an example normalization layer is designed to further improve the performance of the algorithm.Finally,the algorithm is reconstructed by deconvolution.Brightness image.Compared with the existing mainstream image enhancement algorithms,this algorithm can improve the brightness and contrast adaptively and avoid artifacts on image edges.It not only has certain advantages in objective evaluation indexes,but also achieves good results for low illumination images taken in real scenes.An image denoising algorithm of deep convolution neural network based on Inception structure is designed and implemented.Through end-to-end learning,residual learning strategy is adopted to implicitly remove clean images from noisy images and generate noise maps to achieve image denoising.Through residual learning and batch normalization,training speed is improved and performance is improved.The Inception structure is used as the low-level structure of the network.The information of different scales of the image is extracted by multiple convolution kernels.Finally,the image is fused to obtain better features,increase the width of the network and multi-scale invariance.The experimental results show that the algorithm can train a single model of blind image denoising and achieve better results than the specific noise model.A low-resolution image super-resolution restoration algorithm based on antagonistic neural network is designed and implemented.The algorithm is based on generating confrontation network structure.Given the low-resolution and corresponding super-resolution photo sets,the generator in the model is used to generate high-resolution images from low-resolution images,and the sub-pixel convolution layer is used to enhance the image multiplier,while the discriminator is used to distinguish the real image from the generated image.The experimental results show that the super-resolution image generated by the algorithm generator can achieve better results than the traditional method under different magnification.To sum up,aiming at typical image degradation problems such as low illumination,noise and low resolution,this paper carries out the research work of low-quality image enhancement and super-resolution restoration technology based on convolution neural network,which can improve image quality and reconstruct high-quality image from all angles.Relevant technologies will effectively improve the processing effect of various image processing-based application systems for low-quality images,and promote the practical process of various technologies. |
