Abstract: In this paper, a kind of algorithm for all-zero block detection based on Radial Basis Function (RBF) Neural Network (NN) was proposed to improve the accuracy of all-zero block detection algorithm. By analyzing the H.264 encoder features, six effective features were selected, including Sum of Absolute Difference (SAD), Sum of Absolute Transformed Difference (SATD), block type, Rate Distortion Optimization (RDO) cost, Quantization Parameter (QP) and the situation of reference block. Considering the SATD should be used in the Hadamard Transform (HT), to get the relationship of QP and RBF network width parameter through the least square method, the algorithm used two classifiers to separate all-zero blocks from non-all-zero blocks based on the encoding situation of the reference block. This algorithm could improve coding speed over 50% on average while keeping bit rate and video quality almost unchanged. The experimental results show that the proposed algorithm can improve all-zero block detection accuracy effectively and coding efficiency based on NN.

Key words: H.264 encoding, all-zero block detection, Radial Basis Function (RBF) network, Neural Network (NN), Sum of Absolute Difference (SAD), Sum of Absolute Transformed Difference (SATD)

摘要： 针对目前全零块检测算法准确率不高的问题，提出了一种基于径向基函数(RBF)神经网络(NN)的全零块检测算法。通过分析H.264的编码特点，选取了绝对误差和(SAD)、变换绝对差值和(SATD)、编码块类型、率失真优化(RDO)代价、量化系数(QP)、参考块的全零块情况6个特征，考虑了哈达玛变换(HT)中应该使用SATD的情况，采用最小二乘法得到QP与RBF网络宽度参数的关系，根据参考块是否为零，设计了两个分类器来区分全零块与非全零块。在保证图像质量和编码率不变的前提下，平均能提高编码速度50%以上，实验结果表明，利用RBF神经网络很好地提高了全零块检测准确率和编码效率。

关键词: H.264编码, 全零块检测, 径向基函数网络, 神经网络, 绝对误差和, 变换绝对差值和