一种改进的立体图像编码算法

A NEW STEREO IMAGE CODING ALGORITHM BASED UPON SELF-ORGANIZEING NEURAL NETWORK

为了提高左图像的编码效率,提出了一种新的基于自组织神经网络的立体图像编码算法(SOM+VQ+DE),SOM+VQ+DE算法对右图像采用视差估计补偿技术(DE)编码,对左图像则使用基于自组织特征映射算法(SOM)的矢量量化编码来取代传统的JPEG方法,矢量量化与视差估计的残差均使用DCT+霍夫曼进行编码.对立体测试图像Pentagon的实验表明,SOM+VQ+DE算法能够有效地提高左图像的压缩效率:1)在压缩比均为6.5∶1时,SOM+VQ+DE算法的PSNR较JEPG+DE算法提高了2.42 dB;2)在PSNR均为30 dB时,SOM+VQ+DE算法的压缩比改善是JPEG+DE算法的1.8倍.

To improve the coding efficiency of left image, a new stereo image coding algorithm （SOM＋ VQ＋DE） based upon self-organizing feature map （SOM） and disparity estimation （DE） is presented. Vector quantization （VQ） is used to predict left image instead of the general JPEG algorithm. SOM is used for the codebook training. Disparity estimation is used to predict right image as usual. Both the VQ and DE prediction errors are coded by DCT and entropy coding. Experimental results on stereo image Pentagon show that the SOM＋ VQ ＋ DE algorithm effectively improves the coding performance of left image than JPEG ＋ DE algorithm： when compression is the same （6.5： 1）, the improvement in PSNR is 2.42 dB; and when PSNR is the same （30 dB）, the improvement in compression ratio is 1.8 times.