基于改进SOM的纹理合成算法

Synthesizing Texture Using Variant SOM

提出一种基于改进自组织特征映射(SOM)的纹理合成算法,首先给出一种基于纹理的灰度图像的SOM输入向量构造方法,由此得到的输入向量维数是基于TSVQ方法的输入向量的1/3;再对SOM网络的学习规则进行改进使之能正确训练输入向量,最后使用经过SOM网络量化后的向量合成灰度纹理并反向映射到彩色纹理完成纹理合成。实验表明算法合成纹理质量较好,合成速度较TSVQ加速算法有较大提高。

A technique for texture synthesis was outlined using variant Self-Organizing Feature Maps （SOM）. A new method of SOM input vectors construction using corresponding gray image of input texture was proposed in which the vector dimension length is 1/3 of that in previous work of TSVQ method, then the SOM learning rule was extended to train these input vectors. Then the quantized vectors to synthesis a gray texture were used whose pixels could be finally mapped into a color output texture. Experiments show that the quality of output of the method is fine and the synthesis time can be remarkably reduced comparing to TSVQ method.