Research On Image Deblurring Technology Based On Deep Neural Network

Image deblurring,aiming to recover clear images from blurred images,is an important research topic in digital image processing and computer vision,which has a wide application prospect in many fields such as criminal investigation,traffic management,security monitoring,and medical imaging.Motion blur caused by various factors such as camera shake and highspeed motion of objects often superimposes global blur and local non-uniform blur,which brings great challenges to the image deblurring task.Recently,image deblurring methods based on deep neural networks have developed rapidly,and have achieved remarkable results in removing non-uniform motion blur.However,there are still problems such as a large number of network parameters,insufficient feature extraction ability,and poor restoration of local details of the image,and their deblurring performance still has a large space for improvement.Given the inherent defects of existing algorithms,three improved image deblurring methods are proposed in this thesis.More concretely,the contributions of the thesis are summarized as follows:1.an image deblurring network based on channel attention and cross-scale feature fusion is proposed.Given the problems of a large number of network parameters and long inference time in multi-stage image deblurring networks,a single-stage encoder-decoder network structure is designed.In order to extract more effective features and strengthen important feature channels,a residual module with channel attention is introduced.Then a cross-scale feature fusion module is proposed which performs gradual cross-scale feature fusion of shallow features and decoded features of different scales to make full use of the global contour information and local detail information at different scales.Finally,supervision modules are implemented in the different sub-networks of the decoding network to strengthen the constraints on the intermediate decoding features.The experimental results demonstrate that the proposed network can improve the effectiveness of image deblurring effectively while reducing the number of network parameters and the forward inference time.2.An image deblurring method CADNet is proposed in which the main components are deformable convolution layer and fine-grained multi-scale attention residual module(CARes2Net).A deformable convolution encoder-decoder structure is proposed which embeds the deformable convolution into the encoder-decoder network reasonably to improve the resilience to complex blurring deformation.Aiming at the large receptive field and multi-scale feature information required for image deblurring,a fine-grained multi-scale attention residual module(CA-Res2Net)is proposed to represent the multi-scale feature information of images at a more fine-grained level and expand the receptive field of the network.Finally,to enhance the ability to restore high-frequency details of images,a multi-scale frequency domain reconstruction loss function is further introduced on the basis of the multi-scale content reconstruction loss function,which enhances the supervision of restored images in terms of the frequency domain.Experimental results demonstrate that the CADNet method achieves higher image deblurring performance without additional computational overhead,and can restore clearer local details and structures.3.Inspired by the knowledge distillation image dehazing algorithm,an image deblurring algorithm KDDeblur based on knowledge distillation is proposed.KDDeblur consists of two heterogeneous task models,in which the teacher network model is an image reconstruction network based on the self-encoding structure in which the input clear image is reconstructed,and the student network model is an improved image deblurring network based on CADNet.Firstly,the teacher network is trained and optimized.Then,to achieve knowledge distillation from teacher network to student network,a weighted feature imitation loss function is added in the training of the student network,so that the decoding feature layers of the student network at different scales are as close as possible to the clear features in the teacher network.The experimental results demonstrate that the teacher network can effectively improve the deblurring effectiveness of the student network,and the KDDeblur algorithm achieves better deblurring performance while keeping the network parameters relatively small.