摘要
近年来 ,许多学者已经成功地将 Kohonen的自组织特征映射 (SOFM)神经网络应用于矢量量化 (VQ)图象压缩编码 .相对于传统的 L BG算法 ,基本的 SOFM算法的两个主要缺点是计算量大和生成的码书性能较差 ,因此为了改善码书性能 ,对基本的 SOFM算法的权值调整方法作了一些改进 ,同时为了降低计算量 ,又在决定获胜神经元的过程中 ,采用了快速搜索算法 .在将改进的算法用于矢量量化码书设计后 ,并把生成的码书用于图象的压缩编码 .测试结果表明 ,改进的算法使码书设计的计算量得到明显的降低 ,而且码书的性能得到了提高 .相对于基本算法 ,码书设计的计算时间减少了约 75 % .在图象编码中 ,不论是训练集内的图象 ,还是训练集外的图象 ,相对于基本算法 ,编码质量均提高了 0 .80 d B～ 0 .90 d B.
In recent years, many scholars have successfully applied the Kohonens self organizing feature map(SOFM) neural networks to vector quantization image compression encoding. The two main shortcomings of the basic SOFM method are its high computation complexity and its poor codebook quality compared to the conventional LBG algorithm. In order to improve the codebook performance, some modification is made in the weight factor adjustment of the basic SOFM algorithm in this paper. In order to reduce the computation complexity of the basic SOFM algorithm, some fast search methods are used in SOFM iterations during the search for the winning neuron. The proposed algorithm is used to generate vector quantization codebook and the generated codebooks are used for image compression encoding in this paper. Simulation shows that the reduction of computation is substantial and the codebook performance is improved. Compared to the basic SOFM algorithm, the reduction of computation is about 75%. For not only image in the training set but also the image outside the training set, the encoding quality can be improved by 0 80dB～0 90dB compared to the basic SOFM algorithm.
出处
《中国图象图形学报（A辑）》
CSCD
2000年第10期846-850,共5页
Journal of Image and Graphics
关键词
矢量量化
自组织特征映射神经网络
图象压缩
Vector quantization, Self organizing feature map neural network, Image compression
