DecoderNet: A Lightweight Neural Network for Decoder-Side Transform Skip Flag Prediction in Hybrid Video Coding

Abstract

Modern hybrid video codecs such as H.265/HEVC and H.266/VVC support a transform skip mode (TSM) in which certain coding blocks bypass the Discrete Cosine Transform (DCT) entirely and carry raw spatial-domain residuals. A single syntax element, the transform_skip_flag, controls the decode path at the receiver: blocks with this flag set proceed directly to inverse quantization and reconstruction, whereas blocks without it must first undergo the inverse transform. The flag is parsed from the encoded bitstream and, under error-free conditions, governs the decoding pipeline deterministically. However, in lossy or error-prone transmission environments, a corrupted or lost flag forces the decoder to guess, typically causing severe visual artifacts. Furthermore, as video coding research moves toward reduced overhead and implicit syntax, eliminating such flags through reliable prediction is desirable. This paper presents DecoderNet, the first dedicated decoder-side neural architecture for predicting the transform skip decision from locally available contextual information. Our lightweight model, a two-stage binary classifier operating on quantized residual statistics, neighbouring block metadata, and prediction mode information, requires no additional bitstream transmission overhead, adds fewer than 0.3 ms of latency per frame at 1080p, and achieves an average flag-prediction accuracy of 94.7% across standard test sequences. When deployed as an error-resilience module, DecoderNet reduces PSNR degradation under 5% packet loss by up to 3.1 dB compared with conventional concealment methods. These results demonstrate that decoder-side intelligence can meaningfully reduce codec sensitivity to transmission errors and open a credible path toward implicit flag signalling in next-generation video standards.

Keywords