Autoencoder Network Optimization And Application

Autoencoder network is one of the most efficient machine learning algorithms,with very strong nonlinear modeling and robust ability,which can automatically extract useful information,implement it with network weights and performance are selfadapted.As deep learning is prospering,autoencoder is paid more and more attention,and there is more and more applications of autoencoder on image processing.While the Internet is increasingly developing,a large amount of unlabeled and inappropriately labeled images needs to be dealt with.Therefore,more researches on autoencoder have very important theoretical significance and practical value.Traditional autoencoder networks are computationally complex,have poor clarity on reconstruction image,poor subjective perception and bad performance on image clustering with features extracted from them.To solve these problems,autoencoder network is mainly optimized in this paper on network structure and loss measurement:1.The algorithm of autoencoder network based on sub-pixel fully convolutional network is proposed.The proposed algorithm changes the traditional structure of autoencoder of encode layer,decode layer and fully connected layer,fully convolutional layer instead of fully connected layer and sub-pixel network on decode layer.To solve the problem of error accumulation as spatial resolution increasing layer by layer,the proposed algorithm replaces the decode layer with sub-pixel convolutional layer where the data is extracted and brought back to the original data at the last layer.And to maintain the information of spatial coordinates,the proposed algorithm uses fully convolutional layer instead of fully connected layer between the encoder and the decoder.The proposed algorithm has improved 2.93% on PSNR,6.60% on SSIM,47.40% on time.2.The algorithm of autoencoder network based on adversarial MSE(Mean Square Error,MSE)is proposed.The weight calculation network is introduced into traditional autoencoder,with optimized loss function,uses the adversary of two loss measurement of the network to balance high frequency and low frequency components of the image to form better construction image on subjective perception.The experiment on the proposed algorithm shows better performance on subjective perception and has improved 9.8% on PSNR,10.3% on SSIM.