Research On Color JPEG Image Restoration Based On Dual Domain Learning And Convolution Pool

In recent years,deep learning based JPEG compression artifact reduction algo-rithms have made a dramatic breakthrough.However,there are several limitations when applying these algorithms to production.For examples,multi-scale learning models suf-fer from the excessive model size and unnatural restored textures;dual-domain learning models are difficult to be extended to color image restoration and dynamic compression quality restoration;deep models lead to oversized model parameters.For these limita-tions,this thesis studies on deep learning based color image JPEG compression artifact reduction algorithms from aspects including receptive field model,dual-domain learn-ing model and model compressing:First,to solve the excessive model size and unnatural restored textures problems caused by the multi-scale learning models,this thesis proposes a receptive pyramid convolutional network for color image JPEG compression artifact reduction.This al-gorithm firstly proposes a receptive pyramid model for multi-scale feature extraction basing on the enlarging receptive field character of dilated convolution,then proposes a global feature based adaptive color affine transform for color restoration.The exper-imental analysis demonstrates the effectiveness of receptive pyramid model and adap-tive color affine transform,and the reasonableness of parameter settings in proposed algorithms.The proposed algorithm achieves the superior performance with minute parameters.Second,to overcome the limitations of existed dual-domain that being difficult to be extended to color image restoration and dynamic compression quality restora-tion,this thesis proposes an implicit dual-domain convolutional network for color im-age JPEG compression artifact reduction.This algorithm firstly proposes an implicit DCT to predicting the relative quantization losses of channels in YCbCr space,and de-signs corresponding prediction models for different channels,then proposes a statistics prior based pixel labeling method to help network understand the exact positions of pix-els in JPEG image.The experimental results demonstrate that the proposed algorithm makes a significant improvement from state-of-the-art algorithms,and achieves great performance in dynamic compression quality restoration task.Third,to compress dense connection model which is widely used in deep learn-ing based JPEG compression artifact reduction algorithms,this thesis puts forward the concept of convolution pool,and proposes a convolution pool based dense connection model compressing algorithm.When constructing a deep learning model,convolution pool applies all needed parameters at first,then reallocates the parameters to convolu-tion layers by a specific parameter allocating(sharing)mechanism.This algorithm pro-poses a parameter sharing mechanism for dense connection model,which could reduce the depth complexity of dense connection model from O(L2)to O(L).The experimen-tal analysis demonstrates that the proposed algorithm would not lead divergence and overfitting,and is well compatible with dilated convolution.Moreover,the proposed algorithm can significantly enhance parameter efficiency and contribute to building a more compact deep learning model,and well work on existed algorithms.