神經視訊編碼（NeuralVideoCoding）近年在率失真效能與編碼速度上 皆有顯著進展，逐漸具備部署於即時視訊應用的潛力。DCVC-RT透過隱式 時間建模、低解析度潛在表示與可變品質介面，實現兼具高效率與可調整 碼率的神經視訊編碼架構。然而，可變品質能力並不等同於完整的閉迴路 碼率控制（closed-loopratecontrol）。在視訊會議與直播等受限頻寬情境中， 編碼器仍須由目標碼率自動決定控制參數，並依實際輸出位元數持續修正 決策。此外，DCVC-RT的P-frame具有時間特徵傳遞、階層式品質配置與 週期性刷新（refresh）行為，任意的逐幀量化參數（QP）調整可能破壞其 原生時間品質分布與參考相依性，降低率失真效率。為解決上述問題，本 文提出一套針對DCVC-RT的I-frame引導結構感知λ-domain群組式碼率控 制器（GroupRC），在不修改編碼器架構與預訓練權重的前提下建立外部閉 迴路控制機制，並從控制結構、線上模型適應與I/P協作初始化三個層面 進行設計。在控制結構方面，本文配合DCVC-RT原生的八幀階層品質結 構，以P-frame 群組為控制單位，每個群組僅決定一個基礎QP，保留原有 品質階層與時間參考關係，降低碼率控制對率失真效率的破壞。在線上模 型適應方面，本文結合λ-domain碼率模型、遞迴最小平方（RLS）線上更 新與一般／刷新雙狀態分離，避免刷新畫面的異質統計污染估計，使控制 器能依實際編碼結果持續適應內容。在I/P協作初始化方面，本文以內容 感知的I-frame 預測器估計I-QP，並以其操作點資訊初始化第一個P-frame 群組，降低序列起始缺乏觀測的冷啟動偏差。此外，本文提出決策時虛擬 試編碼（virtual response probe），於群組決策點以熵模型估計候選QP的預 期位元數而不實際產生位元流，作為響應觀測、模型驗證與失敗歸因的輔 助工具。實驗於UVG、MCL-JCV與HEVCClassB三個標準測試集上，以固定QP編碼為率失真錨點，並以碼率誤差（BRE）與BD-Rate為評估指 標。實驗結果顯示，相較於直接追求碼率命中的控制規則，GroupRC以部 分碼率命中精度換取率失真效率的顯著改善，有效降低相對固定QP編碼 的率失真代價，取得可解釋的結構性權衡。I-QP初始化在理想估計下展現 明確的協作增益，提升預測精度為進一步改善的直接途徑。虛擬試編碼已 於受控實驗中驗證其量測與診斷能力，為未來內容自適應的線上修正策略 奠定基礎。整體而言，本文方法在保留DCVC-RT原生時間品質結構與率 失真行為的同時進行長序列碼率控制，且額外運算成本相對編碼器極低， 具備即時部署之應用潛力。;Neural video coding has advanced rapidly in rate-distortion performance and codingspeed， makingitincreasinglyviableforreal-timevideoapplications. DCVC RT achieves efficient， bitrate-adjustable neural video coding through implicit tem poral modeling， a low-resolution latent representation， and a variable-quality in terface. However， variable-quality coding is not equivalent to complete closed loop rate control. In bandwidth-constrained scenarios such as video conferencing and live streaming， the encoder must still derive coding parameters from a tar get bitrate and correct subsequent decisions from actual output bits. Moreover， DCVC-RT’s P-frames rely on temporal feature propagation， a hierarchical quality structure， and periodic refresh， so arbitrary frame-level QP adjustment may dis turb this learned structure and degrade rate-distortion efficiency. To address these issues， this thesis proposes an I-frame-informed structure-aware λ-domain group level rate controller (GroupRC) for DCVC-RT， an external closed-loop control mechanism that leaves the encoder architecture and pretrained weights unchanged， designed along three aspects: control structure， online model adaptation， and I/P cooperative initialization. For the control structure， GroupRC adopts the P-frame group， aligned with the native eight-frame hierarchical quality period， as the con trol unit and determines only one base QP per group， preserving the quality hi erarchy and temporal reference relations. For online model adaptation， GroupRC combines a λ-domain rate model with online recursive least squares (RLS) up dates and normal/refresh dual-state separation， so that refresh-frame statistics do not contaminate the estimation and the controller adapts continuously from actual coded bits. For I/P cooperative initialization， a content-adaptive I-frame predic tor estimates the I-QP， whose operating point initializes the first P-frame group and reduces the cold-start bias at the sequence start. In addition， this thesis in troduces a decision-time virtual response probe that estimates the expected bits of candidate QPs with the entropy models at each group decision point without producing any bitstream， serving as an auxiliary tool for response observation， model validation， and failure attribution. Experiments on the UVG， MCL-JCV， and HEVC Class B test sets use fixed-QP encoding as the rate-distortion anchor， with bitrate error (BRE) and BD-Rate as metrics. Results show that， compared with a control rule that directly pursues bitrate accuracy， GroupRC trades part of the bitrate tracking accuracy for a substantial improvement in rate-distortion effi ciency， effectively reducing the rate-distortion penalty relative to fixed-QP encod ing and achieving an interpretable structural trade-off. I-QP initialization exhibits a clear cooperative gain under ideal estimation， and improving the predictor accu racy is a direct path to further gains. The virtual response probe is validated for measurement and diagnosis under controlled experiments， laying the groundwork for future content-adaptive correction strategies. Overall， the proposed method performs long-sequence rate control while preserving DCVC-RT’s native tempo ral quality structure and rate-distortion behavior at low computational overhead， demonstrating its potential for real-time deployment.
