Research On Image Deblurring Algorithm Based On Generative Confrontation Network

As an important carrier for humans to obtain information,images have been widely used in various fields such as computer vision and intelligent transportation.With the rapid development of urban intelligent transportation,the clarity of surveillance images is particularly important.However,during the acquisition and transmission of the image,it will inevitably be interfered by various types of noise.For example,motion blur of the image will occur due to factors such as object movement and camera shake.Therefore,The research of image deblurring technology is very important.In the research of image deblurring algorithms,many scholars have improved the effect of algorithm deblurring by deepening the network and achieved good results.However,this makes the network algorithm more complex,requires a lot of calculations,and consumes a lot of time.Although the deblurring effect of the image has been improved to some extent,it ignores the algorithm's operating speed and has low practical value.Aiming at the problem of image blur removal,this paper proposes a network model to remove motion blur based on generative adversarial networks.The model introduces a multi-scale recursive network as a generator to generate an adversarial network.The generator includes a three-layer network,and overlays multiple layers of residual blocks in each layer to process the fuzzy image from coarse to fine,and passes the discriminator to make a judgment.Secondly,in order to improve the running speed of the algorithm,this paper further proposes an image motion blur removal model based on octave convolution.The model combines a generation adversarial network(GAN)with an octave convolution residual block.The octave convolution residual block is used to construct a multi-scale recursive network as a generator to generate an adversarial network.To the precise processing,the generator uses a multi-scale encoding and decoding structure,which improves the efficiency of the model,and introduces a context module in the last layer of encoding;a time convolution network TCN(Temporal Convolutional Network)passes information across layers,and finally discriminates by a discriminator.Using the octave convolution residual block greatly reduces the amount of model parameters and speeds up the network's processing of the image;the context module uses multi-layered hollow volumes to actively increase the receptive field and better capture multi-scale context information.The main innovations in this article are as follows:1.This paper proposes a network model based on generative adversarial network to remove image motion blur.The model introduces a multi-scale recursive network as a generator for generating adversarial networks,and processes blurred images from coarse to fine to improve the image deblurring effect.The experimental results show that the image after removing the motion blur of the image is clear and the details can be resolved,and the peak signal-to-noise ratio reaches 29.52,which is better than the common model.2.A model of image motion blur removal based on generative adversarial network(GAN)combined with octave convolution residual block and time convolution network(TCN)is further proposed.The model uses an octave convolution residual block as the basic block to reduce network redundancy and speed up the network's processing of the image;the introduction of TCN and the context module increases the receptive field and can better capture multi-scale context information.The experimental results show that the model has achieved good results in terms of visual effects,peak signal-to-noise ratio,and training speed at each step,and has good application prospects.