Flow-Guided ConvLSTM with Quality-Aware Reconstruction for Learned Video Compression

In Learned Video Compression (LVC), contextual information plays a crucial role in reducing temporal redundancy and improving rate-distortion performance. However, most existing approaches rely on a single reference frame or a fixed set of neighboring frames, which limits their ability to exploit long-range temporal dependencies. To overcome this limitation, we propose FGC-QAR (Flow-Guided ConvLSTM with Quality-Aware Reconstruction), a unified framework that combines a Flow-Guided ConvLSTM (FG-ConvLSTM) with a Quality-Aware Reconstruction (QAR) module. FG-ConvLSTM aggregates features across extended temporal spans by explicitly incorporating optical flow as motion guidance. Meanwhile, QAR is implemented using ConvGRU and dynamically modulates its forget and input gates according to the compression quality and content similarity of reference frames. This allows the network to emphasize more reliable temporal cues while suppressing noisy or low-quality information. Extensive experiments demonstrate that the proposed framework consistently improves compression efficiency across multiple datasets and bitrates, highlighting its effectiveness and scalability.