A Processing Network for Transcoding: Bridging Initial and Subsequent Encoding
Miaojun Ni, Mingyi Yang, Hao Wang, Fuzheng Yang · 2025
Video transcoding is essential for multimedia processing as it enhances transmission efficiency, supports a variety of devices, and improves the user's experience. However, the output from the initial encoder is often unfriendly to subsequent transcoding. Existing transcoding optimization methods focus either concentrate on the initial encoding or the subsequent transcoding, neglecting the interplay between the two, even though both encoders significantly impact the overall transcoding process. In this work, we propose a processing network that bridges the initial encoding and subsequent transcoding, enabling a joint optimization of the transcoding process. For the initial encoder, considering the areas with lower residuals typically have smaller quantization losses, whereas areas with higher residuals do not, we employ residuals to guide the network in restoring compression distortion. In parallel, for the joint optimization of the subsequent encoder and the processing network, considering areas with large quantization losses typically indicate that the original region's distribution is either unsuitable for encoding or has complex textures, we have developed a corresponding mask in the DCT domain, and employ the quantified loss distribution from the subsequent encoder to fine-tune the loss training of the processing network. See Figure 1 for more details. Experiments show substantial enhancements in transcoding performance when transitioning from H.264 to H.265.