Research On Dynamic Scene Deblurring Based On Convolutional Neural Network

Due to the relative motion of the device and the target,images obtained under dynamic scenes often have non-uniform motion blur.At present,traditional image deblurring algorithms use iterative optimization to reconstruct clear images.The quality of reconstructed images under complex motion blur is low,and the real-time performance is poor.In order to effectively remove the blur in dynamic scenes,the paper uses a convolutional neural network to solve the problem of deblurring dynamic scenes.The specific research contents are as follows:Firstly,the paper proposes a multi-scale and multi-stage residual attention network dynamic scene deblurring method.The method introduces progressive ideas,while the network gradually recovers latent images from blurry images by stacking encoder-decoder modules,and obtains a large overall receptive field.In order to obtain multi-scale information of the image,the method follows the coarse-to-fine and recursive ideas,using different sizes of the image to jointly train the network.The network uses the channel attention mechanism to rescale image features and improve its own representation ability.The experiments show that the proposed method has strong generalization ability.Secondly,in order to reduce the complexity of mapping and improve the efficiency of network learning,the paper proposes a multi-scale dynamic scene deblurring method based on wavelet transform.The method uses the multi-scale input strategy to obtain the multi-scale information of the original image,while stacking the Residual-Inception module to extract the multi-scale deep features of images.The network uses the dual attention mechanism to improve its expressive ability from the perspective of channels and space,respectively.Experiments show that the proposed method has better performance on public data sets.Finally,in order to further improve the network performance,the paper proposes a deep multilevel wavelet dynamic scene deblurring method.The wavelet transform is embedded in the encoding-decoding structure,which enhances the sparsity of image features while increasing the receptive field.The method uses multi-scale expansion dense blocks to extract multi-scale deep features,introduces feature fusion blocks to adaptively fuse features between encoding and decoding,and uses the spatial domain reconstruction module to further improve the quality of reconstructed images in the spatial domain.The experiments show that the proposed method has obvious performance improvement and reconstructs more image details.