Research On Single Image Deblurring Based On Convolutional Neural Network

With the development of artificial intelligence,machine vision is widely used in the fields of video surveillance,medical treatment,diagnosis,and autonomous driving.However,due to the uneven quality of the camera’s light-sensitive unit,the lack of professionalism of the shooter,the harsh shooting environment,and the movement of the camera or the object,the captured images contain obvious blur.Therefore,adopting image processing technology to restore clear images from blurred images is of great practical application.In this paper,convolutional neural networks are used as a research tool to focus on the single image non-uniform motion blurring task.We achieve efficient image deblurring performance by designing a reasonably network.The specific research contents are summarized as follows:1.Existing image deblurring convolutional neural networks mainly use multi-scale iterative or network recurrent strategies,which will lead to excessive amount of parameters and high computational cost.Therefore,a lightweight multi-scale fusion coding deblurring network(MFC-Net)is proposed in this paper.The network fuses the image multi-scale information into the encoder of the single-scale deblurring network,which effectively fuses the image multi-scale features,and speeds up the training with super deblurring performance.In addition,for local blur inconsistency in non-uniform blurred images,we propose a region attention module to enhance the network’s attention to local blur by dynamically weighting local features into feature vectors of different perceptual fields.Meanwhile,the feature fusion module is designed to enhance the correlation between features at different scales.Finally,we adopt the generative adversarial network(WGAN)to recover more natural and realistic images.Numerous experimental results show that our proposed network outperforms existing methods in terms of performance,and speed.2.In order to make full use of the information of different scale features,we propose a multi-input W-network(MI-WNet)based on augmented contextual information.Firstly,the single-scale U-shaped network is transformed into a W-shaped network by retaining the image multi-scale input coding fusion on the above MFC-Net base framework.By this way,the features can be more effectively fused and the training will be more easier.In addition,to address the limitations of the network in mining feature information,we design a dense context augmented module between the encoder and decoder,which can effectively expand the sensory field and obtain multi-scale information.We further use a dense connection to reuse the feature effectively and make the features more compact.The experimental results show that the proposed MI-WNet has strong generalization ability and superior deblurring performance.3.Image deblurring deployment and visualization.In this paper,the above MFC-Net and MI-WNet network models are deployed online to build a real-time image deblurring system.The online visualization of the deblurring effect is achieved,which helps to analyze the actual deblurring effect visually.At the same time,the ability to replace the network model online helps to justify the image deblurring network.