Image Deblurring Based On Convolutional Neural Network

When taking an image in an actual scene,the captured image contains blurry conditions that are not conducive to obtaining information due to the movement of objects and camera shake.The purpose of image deblurring is to restore the blurred image to clarity through technical means,which is a popular research direction in the underlying tasks of computer vision.The existing image deblurring methods have solved the problem of image blur well,but there are still certain deficiencies in the extraction of image feature information,the inference speed of processing blurred images,the processing of image local texture information,and the production of blurred image datasets.In recent years,deep learning has developed rapidly,and deep learning image processing methods based on convolutional neural networks(Convolutional Neural Network)have been widely used in various fields.On the one hand,processing blurred images in an end-to-end form improves the quality of restored images,on the other hand,speed up the processing of blurred images achieving the effect of real-time restoration of clear images.The main innovative ideas of this paper are as follows:1.An image deblurring based on multi-scale feature network is proposed,which combines multiple sub-networks,and the output of the previous stage is used as the input of the subsequent stage.Aiming at the different input image resolutions in different stages,a multi-scale feature extraction module is proposed.The different stages are connected by a cross-stage attention module to speed up the inference time and achieve better results.2.An image deblurring method based on augmented multi-scale features and attention is designed,which reduces the number of stages of the network model.On the basis of the previous enhanced multi-scale feature extraction module,Introducing the normalization function to achieve the purpose of fully obtaining the effective feature information of the image.An attention module is added between the decoding modules to help improve the quality of the restored image.This method can fully extract the features of the image,and filter to obtain effective semantic information,which has a better deblurring effect.3.An image deblurring network based on improved multi-scale and attention features is proposed.For the improvement of the previous model,the cross-stage feature fusion module is removed,and the original image features are replaced with attention feature maps transmitted to the next stage.At the same time,the attention mechanism used in the encoder-decoder is improved,the information fusion between feature maps of different scales is used,and the weight value of important information is amplified through the attention mechanism to improve the utilization efficiency of image features.In general,the single-image deblurring algorithm proposed in this paper has been tested on multiple blurred image datasets,and compared with other existing deblurring methods based on convolutional neural networks.The quality of the clear images recovered by the proposed methods in this article are higher,and the details of the images are processed better.