Deep Learning based Fast Mode Decision in HEVC Intra Prediction using Region Wise Feature Classification
View/ Open
Date
2018-12-07Author
Kuanar, Shiba Prasad
0000-0002-8670-6957
Metadata
Show full item recordAbstract
The High Efficiency Video Coding (HEVC) standard has achieved best coding efficiency as compared to previous H.264/AVC standard. But the computational time of HEVC encoder has increased mainly because of the hierarchical quad-tree based structure, recursive search for finding the best coding units, and the exhaustive prediction search up-to 35 modes. These advances improve the coding efficiency, but result into a very high computational complexity. Furthermore selecting the optimal modes among all prediction modes are necessary for the subsequent rate distortion optimization process.Therefore we propose a convolutional neural network (CNN) based algorithm which learns the region wise image features and performs a classification job. These classification results are later used in the encoder downstream systems for finding the optimal coding units in each of the tree blocks, and subsequently reduce the number of prediction modes. For our model training, we gathered a new data-set which includes diverse images for the better generalization of our results. The experimental results show that our proposed learning based algorithm reduces the encoder time up to 66.15 % with a minimal Bjontegaard Delta Bit Rate (BD-BR) loss of 1.34 % over the state-of-the-art machine learning approaches. Furthermore our method also reduces the mode selection by 45.91 % with respect to the HEVC baseline.