采用轻量级卷积神经网络的 H.266/通用视频编码跨分量预测

To improve the accuracy of intra chroma prediction in H. 266/versatile video coding (WC) > a cross-component prediction method based on lightweight convolutional neural network was proposed in this paper. The luma modulc and chroma modulc wcrc dcsigncd to extract features from luma and chroma reference samplcs, and the attention modulc was designed to leverage the attention mechanism to construet the spatial correlation between the current luma reference samplcs and the boundary luma reference samplcs. Finally, the attention mask was applied to the boundary chroma reference samples to generate chroma prediction value. To reduce the encoding and decoding complexity, the feature fusion and prediction in the network were achieved in two dimensions, the existing training strategy with shared Parameters to handle variable block sizes was improved, and slimmable convolutions were introduced to adjust network Parameters according to diffcrcnt block sizcs. The cxperimental results show that the proposed algorithm achieved 0. 30%/2. 46%/2. 25% BD-rate reduetion on the Y/Cb/Cr component, respectively, compared with the H. 266/VVC test model VTM18. 0. Compared with other convolutional neural networks-based cross-component prediction methods, the proposed method effectively reduced the network Parameters and inference complcxity, saving 10% encoding time and 19% decoding time.