Research On Convolutional Neural Network Based Distortion Prediction For Image Compression

In signal compression,distortion information is significant for rate distortion optimization.In this paper,we propose a convolutional neural network(CNN)to predict distortion information for H.265/HEVC.With the strong representation power of CNN,structural similarity(SSIM)maps can be predicted directly in an end-to-end,pixel-to-pixel way.Different from traditional CNNs which focus on learning one-to-one mappings from input to output,we show that our CNN model can predict SSIM maps conditioned on quantization parameters(QPs),and realize one-to-many mappings.To construct our CNN network,QP labels are designed as conditions to feed the CNN model.Four residual blocks are used for better feature extraction.We also apply symmetrical network architecture and multi-level feature fusion method to ensure our network can utilize both high-level semantic features and low-level structure features.With an exactly symmetrical structure,our network can fuse the features of symmetrical layers easily since they have the same feature size.Symmetrical structure also ensures the prediction accuracy because we don't have to scale the feature to the target size for feature fusion,which will change the original value because of interpolation.The experiments on MS COCO database demonstrate the effectiveness of our CNNbased method for SSIM prediction.We show that SSIM maps can be predicted conditioned on QP labels without training models separately for each QP.Our multiple QP prediction model even outperforms the single QP prediction model,which demonstrates the robustness and effectiveness of the proposed CNN model.In addition,the proposed QP label enables our CNN model to train a unified model for all QPs,to solve time-consuming problem in training process.A unified model is also more suitable in practical applications with QP label as the condition to decide the outputs of distortion maps.We can also seek different input labels that can be the conditions of our CNN model to realize more generalized distortion prediction in one unified CNN model.In this way,our model can not only predict SSIM maps conditioned on QP labels,but also conditioned on any other parameters related to the distortion information.