高通量图像编码中的端到端量化参数优化方法

End-to-end quantization parameter optimization for high-throughput image coding in real time

高通量的图像传输可以获得更多的图像细节信息.在传输带宽受限和图像间时域相关性很低的条件下,图像编码的输出受到实时性和码率两方面的约束.有损图像编码的量化参数对输出码率和图像质量都有非常重要的影响.该文不同于基于图像复杂性特征的量化参数确定方法,提出了端到端的卷积神经网络深度模型、直接从图像预测最佳量化系数的方法.考虑编码实时性和算法泛化能力,在Inria aerial image labeling dataset数据集上训练,得到了优化的网络结构.实验结果表明,该文提出的端到端量化参数预测方法相比较相位一致性参数、SATD、图像信息熵等图像特征参数方法,码率预测准确度相较线性回归方法提高了10.31%,相较多层感知器方法提高了8.57%.

In recent years,high throughput image transmission has become widely used due to its ability to obtain graphic details.Under the circumstances of limited bandwidth and low temporal correlation between images,image coding strategies should satisfy with the real-time requirement and bandwidth available.In lossy compression algorithms,the quantization parameter(QP)affects greatly both output bitrate and image quality.Different from QP optimization strategies based on numerical image features such as the sum of absolute transformed difference(SATD),phase congruency(PC),structural similarity(SSIM),a CNN-based end-to-end rate control solution was proposed,which predict the optimal QP directly from images.Trained on Inria Aerial Image Labeling Dataset,the refined rate control model is robust under real-time scenes.Experimental results show that the proposed end-to-end rate control method can achieve the target bitrates by 10.31%bit rate accuracy(BRA)more accurately than the original rate control algorithms based on numerical image features.The proposed method also achieves 8.57%BRA gain compared to multilayer perceptron(MLP)method.