模糊强化学习型的图像矢量量化算法

A New Method of Image Vector Quantization Using Fuzzy Reinforced Learning

本文给出了一种新的图像矢量量化码书的优化设计方法.传统矢量量化方法只考虑了码字与训练矢量之间的吸引影响,所以约束了最优解的寻解空间.本文提出了一种新的学习机理———模糊强化学习机制,该机制在传统的吸引因子基础上,引入新的排斥因子,极大地释放了吸引因子对最优解的寻解空间的约束.新的模糊强化学习机制没有采用引入随机扰动的方法来避免陷入局部最优码书,而是通过吸引因子和排斥因子的合力作用,较准确地确定了每个码字的最佳移动方向,从而使整体码书向全局最优解靠近.实验结果表明,基于模糊强化学习机制的矢量量化算法始终稳定地取得显著优于模糊K-means算法的性能,较好地解决了矢量量化中的码书设计容易陷入局部极小和初始码书影响优化结果的问题.

This paper presents a new method toward the design of optimized codebooks by vector quantization （VQ）. The conventional VQ techniques is easy to converge in a local optimum codebook, which is near to the initial codebook because only the attraction of each training vector and codevector is considered in these techniques, A strategy of fuzzy reinforced learning （FRL） is proposed where not only the attractive factor but repulsive factor are integrated into each iteration of FRL. Codevectors move intelligently and intentionally toward an improved optimum codebook design. Within each iteration of FRL, the size and the direction of the movement of each codevector is determined by the overall pairwise competition between the attractive factor of each training vector and the repulsive factor of its corresponding winning codevector. This new fuzzy reinforced learning vector quantization （FRLVQ） is distinct from some improved VQ techniques in which only randomly generated perturbation is applied to the codebook at each iteration. Experiment results have demonstrated that FRLVQ reduces not only its tendency of becoming trapped in a local optimum but its dependence in the selection of the initial codebook. The consistently superior results are obtained by FRLVQ in comparison with the behavior of well-known fuzzy K-means algorithm.