Scene character recognition is challenging owing to various noise and distortions. In addition, Japanese character recognition requires a large number of training data since thousands of character classes exist in the language. Some researches proposed training data augmentation techniques using synthetic scene character data (SSD)to compensate for the shortage of training data. We proposed multi-scale scheme using multiple dataset consisting of SSD in our previous work to improve recognition accuracy. For further improvement of the scheme, we then proposed Random Filter as SSD generator. This paper enhances the effectiveness of the multi-scale scheme and Random Filter by using larger dataset, JPSC1400, which we have developed by collecting Japanese real scene characters. Experimental results show that the accuracy has been improved from 61.5% to 65.5 % by newly introduced an affine transformation to the SSD generation.