一种基于联想记忆系统实现图像压缩的新方法

Novel method of image compression based on associative memory system

针对传统神经网络用于图像压缩时存在的训练时间长、泛化能力弱等问题 ,提出一种基于联想记忆型神经网络的图像压缩新方法 .利用牛顿前向插值多项式构建联想记忆系统 ,对图像数据进行建模 .首先将图像数据分为多个数据块 ,然后利用数据块对联想记忆系统进行训练 ,训练结束后得到该数据块的特征数据 ,特征数据的数量小于原始数据块 ,且数值大多在零附近 .最后对所有数据块的特征数据重新排序 ,进行熵编码 ,从而实现图像数据的压缩 .实验结果表明该方法是可行的和有效的 ,相比传统神经网络 ,联想记忆系统无需预先训练 ,不依赖训练集数据和初始值 ,可以实时编码 .

To study the traditional neural networks which were featured as slow convergence and poor generalized capacity in image compression, a novel method of image compression based on associative-memory-system neural network was proposed. Associative memory system was constructed by Newton's forward interpolation polynomial, and was used to establish model for image data. First, image data were divided into many blocks. And then each block was utilized to train associative memory system and characteristic data can be abstracted after training. Characteristic data number was less than original data block, most of characteristic data were limited to a range near to zero. Finally, all characteristic data were ranged by special order and entropy encode was exploited to code these characteristic data. Experiments show that the method is effective for image compression. Compared with previous neural networks used in image compression, this method is free of training in advance and converges more quickly.