Research On Image Deblurring Based On Feature-Enhanced Generative Adversarial Networks

Image is an important source for people to obtain and exchange information from the outside world.However,there are many factors in the process of image acquisition,such as object motion and camera shake.The image is blurred and distorted,which seriously affects the comfort and satisfaction of image viewing.It brings great obstacles to the subsequent image analysis and understanding.Therefore,it is extremely important to research on deblurring technology to recover clear and impressive images from distorted images.In recent years,most methods based on deep learning improve the deblurring effect by increasing network complexity.Although the effect is improved,it increases the running time of the algorithm.Sometimes artifacts are inevitably introduced and may not be able to handle complex motion blur.Aiming at the problem of moving image deblurring,this paper uses generative adversarial networks as the research base point,and mainly focuses on the deblurring of natural images as the research goal.A deblurring method based on feature-enhanced generative adversarial networks is proposed,which aims to improve the performance of the model by enhancing image feature extraction.The main work is as follows:(1)This paper analyzes the basic principles and characteristics of existing deblurring methods in detail and concludes that the deblurring effect has room for improvement in feature extraction.Therefore,this paper mainly conducts in-depth research on image deblurring feature extraction technology.(2)In order to enable the network to capture more complex features,a feature extraction module based on the dual-domain idea and a spatially continuous learnable activation function is proposed.On the one hand,in view of the shortcoming that purely performing feature extraction in the spatial domain will produce excessive information loss,the feature extraction is performed through the complementary spatial and frequency domain dual-domain convolution to improve the network performance in the feature extraction part.On the other hand,the joint spatial correlation realizes a non-linear learnable activation function,which is different from the traditional pixel-by-pixel activation unit.In the case of the network has same performance,the activation function not only increases the density of non-linear feature mapping,but also can have fewer network layers.The experimental results show that this method is very helpful to the improvement of network model performance and it is an effective feature extraction method.(3)In order to further improve the effect of image deblurring,combining the above-mentioned research strategies,by introducing the idea of multi-level residual learning and sub-pixel convolutional layer,a deblurring research method based on feature-enhanced generation confrontation network is proposed.We have conducted a lot of experiments and collected a set of real-world blurred image data sets using different shooting equipment to verify the effectiveness of the proposed method for real-world image deblurring.Compared with other methods,the method in this article can recover clearer and more detailed images from synthetic and real-world blurred images.Aiming at the direction of image deblurring,the paper focuses on the problem of feature extraction in the image processing process and proposes a feature extraction strategy based on dual-domain convolution and a spatially continuous learnable activation function.Finally,this paper proposes a image deblurring research method which combined with the idea of multi-level residual learning and pixel recombination technology integrates the generator structure in the generative confrontation network.Through a large number of experimental quantitative evaluation and analysis,the PSNR value on the GoPro test data set reached 31.13,and the SSIM value was 0.9420.The results show that the method in this paper has achieved good performance.