A Removing Motion Blur Model To Facial Images Based On Facial Motion Reordering

High-resolution facial images play an important role in identity recognition and other fields.During image acquisition,some factors such as camera movement and changes in posture can cause motion blur in facial images.Some methods based on deep learning are an effective technological approach to remove motion blur of facial images.However,some existing methods suffer from the problems of large number of parameters and inaccurate extraction of time sequences from blurred images.In order to solve the above problems,a removing motion blur model to facial images based on facial motion reordering is constructed in this thesis.Some contributions of this thesis are as follows:1.Aiming at the problems of large number of parameters and complex calculation of some existing facial image removing motion blur methods,a removing motion blur model to facial images based on multivariate generative adversarial network is constructed,which consists of a generator and a discriminator.In the generator,the traditional generator architecture is changed to a multiple-input and multiple-output U-shaped network architecture,which consists of a feature extraction module,an encoder module,an asymmetric feature fusion module and a decoder module,which can receive and fuse multi-scale information flow,reduce the amount of parameters and reduce the computational complexity.In the discriminator,the traditional discriminator architecture is changed to a codec network architecture,the network architecture can provide more information feedback for the generator and further improve the deblurring performance of the generator.The test results of the 300 VW dataset show that the peak signalto-noise ratio,structural similarity and parameter quantity of the model are 34.0451 d B,0.9356 and 6.806 M,respectively.2.To address the issue of inaccurate extraction of time sequences from blurred images in the removing motion blur model to facial images based on multivariate generative adversarial network,this paper proposes a removing motion blur model to facial images based on facial motion reordering.The proposed model introduces a mapping network and a controlled adaptive block in the generator.The mapping network improves deblurring performance by changing continuous control factors,the control adaptive block extracts and fuses features from blurred images and control factors.An auxiliary regression module has been introduced in the discriminator,which provides regression values for the estimated control factor for better optimization of the generator.The model can extract the accurate time order in the motion blur image,which solves the problem that the accurate time order in the blurred image cannot be extracted in the face motion blur.The test results in the 300 VW dataset show that the peak signal-to-noise ratio and structural similarity of the model are 34.2216 d B and 0.9361,respectively,which are 0.54% and 0.11% higher than that of the MIMO model,and the parameter quantity is 6.808 M,which is 0.029% less than that of the MIMO model.