Frame-level adaptive variable length coding for MPEG-4 AVC/H.264

In conventional video coding, transforms such as DCT have been playing a crucial role by providing good data compaction for further compression with quantization and entropy coding. Traditionally, transformed coefficients after quantization are represented with nonzero coefficients and their locations (or patterns) in a block. The compression of coefficient patterns was often done with run-length coding, which is also the case in MPEG-4 AVC/H.264. In this paper, we propose a new method to compress the coefficient patterns adaptively on the frame-basis. In the proposed method, we first categorize all possible coefficient patterns (i.e., 2¹⁶ -1 = 65535 for 4 × 4 transform in AVC) into 16 different classes based on the number of nonzero coefficients in a block. The proposed method exploits the fact that the coefficient patterns have strong correlation depending on the video sequence and quantization parameter. The probability of each coefficient pattern is updated after encoding each frame, which is then used to generate a new variable length code of the pattern for the next frame. Experimental results show that the proposed method can provide up to 9.4% bit saving over MPEG-4 AVC/H.264 baseline profile encoder.