在現今的社會，我們對於解析度的要求越來越高，為了因應我們所需高解析度的影像，多功能影像編碼（VVC）能比上一代的視訊編碼(HEVC)高出兩倍的壓縮率，這是因為在HEVC的壓縮技術中，使用了編碼單元、預測單元、轉換單元以及量化等方式。在網路傳輸方面，碼率控制是為了使傳輸的影像在特定的通道容量下有較低的失真量以及較好的效能，本論文採用R-λ model控制在low-delay 配置下CTU級的比特率，對於畫面間預測，發現部分CTU的位元錯誤率和像素域的變異數以及運動向量的變異數呈正相關，在R-λ model下的碼率控制算法非常依賴λ值的精確度，達成精確的目標比特率，在本論文中，為了能使畫面間預測有更精確的碼率控制方法，利用統計決定像素域變異數以及運動向量變異數特定閥值，若大於閥值，則引用卷積神經網路來預測編碼樹單元的參數，實驗結果表明，基於卷積神經網路的方法，相較於多功能影像編碼中VTM 20.0的碼率控制方法的部分，在位元錯誤率方面降低了0.122%。;In today′s society, our demand for resolution is increasing. To meet the requirements for high-resolution images, Versatile Video Coding (VVC) can achieve a compression rate twice as high as the previous generation video coding standard, High Efficiency Video Coding (HEVC). This improvement is attributed to the use of coding units, prediction units, transform units, and quantization in HEVC′s compression techniques.
In terms of network transmission, bitrate control aims to achieve lower distortion and better performance of transmitted images within a specific channel capacity. This paper adopts an R-λ model to control the bit rate at the Coding Tree Unit (CTU) level under low-delay configuration. For inter-frame prediction, it is observed that the bit error rate of some CTUs is positively correlated with pixel domain variance and motion vector variance. The bitrate control algorithm under the R-λ model heavily relies on the accuracy of the λ value to achieve precise target bit rates.
In this paper, for more accurate bitrate control in inter-frame prediction, a statistical approach is used to determine specific thresholds for pixel domain variance and motion vector variance. If these variances exceed the threshold, a Convolutional Neural Network (CNN) is invoked to predict the parameters of the coding tree unit. Experimental results demonstrate that the CNN-based approach, compared to the bitrate control method in VTM 20.0 of VVC, reduces the bit error rate by 0.122%.
