Abstract: The new generation video coding standard H. 265/HEVC employs in-loop filter，which includes de-bloc- king (DBF) and sample adaptive offset filter (SAO)，to remove the blocking artifacts and reduce the distortions of reconstructed video frames． Both of DBF and SAO originated from signal processing theory，and the corresponding algorithms and parameters are designed and set manually． Although the computational complexity is relatively low， such filters may not deal with different kinds of contents well enough as the natural videos are much more complex． This paper formulates the loop-filter problem in video coding as an end-to-end regression problem，which can be solved by deep neural network． The relationship between reconstructed frames and original frames are mapped au- tomatically and as a result，the differences between them are minimized． The proposed Multi-Feature based Incre- mental Learning Network (MFILNet) includes 35 layers． The integrated network adopts global residual learning strategy and cascades several Feature Incremental Learning Blocks (FIBs) to extract features of different levels． Consequently，useful features are finally extracted，selected and enhanced to improve the perceptual ability of the network． Within each FIB，variable convolutional kernels are adopted． Inspirited by DenseNet，features from dif- ferent layers are fused，thus to facilitate information flow among layers． Experimental results show that with the scheme of combining density and sparsity，learning capability and generalization capability of the proposed network are boosted tremendously． Both objective and subjective quality of the video compressed frames is improved signifi- cantly． Consequently，the proposed network model is used to substitute the DBF and SAO in H. 265/HEVC． Up to 11. 2% and averaged 6. 32% BD-rate reduction is obtained． The model is also used after the DBF and SAO， 5. 24% BD-rate saving can be obtained in average．

Key words:  H. 265/HEVC, in-loop filter, convolutional neural network, incremental learning