利用擾動法校正TMN8量化參數

Abstract

在H.263標準中使用TMN8速率控制機制，在TMN8的macroblock層速率控制演算法中，理論上是希望總位元數可以按照每個macroblock的變異數 與失真權重 的乘積成正比來分配，其做法是利用二個參數：斜率參數K與標頭參數C求出最佳量化參數Q，利用求出的Q值對DCT係數做量化，希望得到的位元數會等於理論值。但理論值跟實際壓縮得到的位元數還是會有差異，所以需要有一種更新的機制來控制真正壓縮得到的總位元數，TMN8的更新方法是在每一次壓縮完一個macroblock便重新計算剩下的總位元數再進行分配，而且每次都會利用統計方法來平均實際值與預測的K與C值。如此做法雖然可以得到相當接近通道速率的總位元數，但還是存在一些問題而導致某些frame的位元分配並沒有正比於變異數 與失真權重 的乘積。本論文則是提出另一套更新K與C的方法，由於C是標頭參數，用來預測標頭位元，而有些標頭檔是在做速率控制前便已知，其他幾個標頭檔也可以用統計的方法得到一個相近的平均值，如此可以不必像TMN8中將C設為完全未知來預測，本方法同樣也會有一套更新標頭參數的機制，相信會比TMN8得到更準確的C值。至於K的部份，則是求出每個macroblock理論上應分配到的位元數與實際壓縮得到的位元數之差異，使原本該差異值應該反應在總位元數轉變成反應在斜率參數K，除此之外，本方法會給予每一個frame適合的K的初始值，有別於TMN8延用前一個frame的最後一個K值，如此亦可較接近理論的位元分配值。 在第四章的模擬結果可以看出，本方法有效改善位元分配不均的情形，除了可以降低理論值與實際值的差異之外，還可以得到較好的PSNR值，由於主要還是延用TMN8速率控制演算法，所以總位元數還是會接近通道速率。

The research is based on TMN8 rate control method in H.263. We know that, theoretically, the bit allocation by TMN8 for each macroblock is proportional to the product of the deviation value and the distortion weighting . The way to find the optimized quantization value Q is to predict two parameters, K and C. Although the approach to find K and C in TMN8 results in constant frame rate, the bit allocation in each macroblock usually is not the same as the theoretical value for some reasons. Now we want to find a new approach to get K and C. By the new K and C, we will get a new optimized quantization value Q that can be used to encode the macroblock and will make the bit allocation close to the theoretical value. The parameter C is called overhead rate. In H.263 syntax, some header fields are known before rate control. We can get a new C value by finding these known fields and predicting the others. As mentioned before, the theoretical bit allocated value is usually not the same as the actual bit rate. We find the difference between the theoretical bit allocated value and the actual bit rate. According to the difference value, we update the parameter K by some algorithms. By the simulation results, we can see that the new approach to update K and C will let the actual bit rate closer to the theoretical bit allocated value. Because the new approach is still based on TMN8, we still can have constant frame rate. The benefits are getting greater PSNR and reducing skip blocks in a frame.