Model based rate control for predictive video encoder
View Patent ↗A rate control algorithm for a predictive video encoder is provided to accurately control the coding rate while requiring a small complexity. The algorithm assumes the video encoder to consist of two processors, pre-processor and main processor, and the rate control located in main processor uses prediction cost and residual cost for each picture that are computed in pre-processor. Prediction cost represents the estimated number of bits needed to generate an error image, and residual cost describes the difficulty of encoding the resultant error image. Accurate and simple rate control is made possible by using a new rate model that adapts to the local characteristic of a video sequence, and by targeting to maintain the quantization parameter as constant as possible in a given set of pictures, GOP, instead of the conventional two step approach.
1. A method of performing rate control in a predictive video encoder, comprising:
classifying pictures in a video sequence into respective categories;
estimating a first quantization parameter;
calculating a respective prediction cost and residual cost for each of the pictures in the video sequence based on the estimated quantization parameter;
initializing rate control data for each category;
computing a respective picture cost for each picture using the respective prediction cost and residual cost, and the rate control data for the respective category;
performing bit allocation and quantization parameter selection for a picture selected from the video sequence using the respective picture cost and the respective rate control data for all the pictures in a respective group of pictures to which the selected picture belongs;
encoding the selected picture;
updating the rate control data for the respective category; and
performing bit allocation and quantization parameter selection for a second picture selected from the video sequence using the respective picture cost and the respective rate control data for all the pictures in a respective group of pictures to which the second picture belongs.
2. The method of claim 1 , wherein the picture cost is a weighted sum of prediction cost and residual cost.
3. The method of claim 2 , wherein the prediction cost is an estimate of the number of bits needed to generate an error image, and the residual cost describes the difficulty of encoding the resultant error image.
4. The method of claim 1 , wherein classifying pictures in the video sequence comprises:
detecting scene changes in the video sequence; and
setting a group of picture (GOP) size based on the detected scene changes.
5. The method of claim 1 , further comprising:
classifying each picture in a GOP, wherein the categories characterize different coding schemes and availability of certain reference pictures for encoding pictures in each category.
6. The method of claim 5 , wherein each category is characterized by the average quantization parameter used to encode the pictures in a specific category, the average prediction cost for the category, the average residual cost for the category, the relationship between the prediction cost and the actual number of bits used, the relationship between the residual cost and the actual number of bits used to encode the picture, and the total picture cost.
7. The method of claim 1 , wherein the rate control data comprises:
data representing the quantization parameter used to encode the pictures in a category;
data representing average prediction costs;
data representing average residual costs;
data representing the relationship between the prediction cost and an actual number of bits used in prediction;
data representing the relationship between the residual cost and an actual number of bits used to encode transform coefficients; and
data representing the actual number of bits used to encode a picture.
8. The method of claim 1 , wherein updating the control data comprises:
calculating a quantization parameter for the GOP as a weighted average of quantization parameters stored in the control data;
calculating the picture cost for the pictures in the GOP;
calculating the number of bits used to encode the ith picture; and
calculating a quantization parameter that generates a number of bits closest to the target number of bits.
9. A predictive video encoder, comprising:
a pre-processor configured to
classify pictures in a video sequence into respective categories,
estimate a first ciuantization parameter, and
calculate a respective prediction cost and residual cost for each of the pictures in the video sequence based on the estimated quantization parameter; and
a main processor configured to initialize rate control data for each category,
compute a respective picture cost for each picture using the respective prediction cost and residual cost, and the rate control data for the respective category,
perform bit allocation and ciuantization parameter selection for a picture selected from the video sequence using the respective picture cost and the respective rate control data for all the pictures in a respective group of pictures to which the selected picture belongs,
encode the selected picture,
update the rate control data for the respective category, and
perform bit allocation and ciuantization parameter selection for a second picture selected from the video seciuence using the respective picture cost and the respective rate control data for all the pictures in a respective group of pictures to which the second picture belongs.
10. The encoder of claim 9 , wherein the picture cost is a weighted sum of prediction cost and residual cost.
11. The encoder of claim 10 , wherein prediction cost is an estimate of the number of bits needed to generate an error image, and residual cost describes the difficulty of encoding the resultant error image.
12. The encoder of claim 9 , wherein the pre-processor is further configured to:
classify scenes by determining scene changes, flash pictures, and still pictures in the video sequence; and
set the GOP size based on the classification of scenes.