It is convenient for the visually handicapped to recognize specific words printed on the surfaces of commercial products. We propose a method that recognizes specific words by matching them to a dataset generated from font data. We extracted approximately 9,000 kinds of character images in each of two font types (MS Mincho and MS Gothic), including Chinese characters (kanji), and the Japanese systems hiragana and katakana. The characters are classified into 23 classes; 21 important characters to be recognized (classes 1-21), other characters (class 22), and no characters (class 23). To diversify the kinds of character images, we altered their color and added blur. The final dataset includes over one million images. Next, a convolutional neural network (CNN) was trained with this dataset. In the test phase, character candidates were extracted from images of commercial products by binarization and edge extraction. For character recognition, the character string was converted into meaningful words by thresholding the distances between characters. In an experimental verification using the intersection over union criterion, the word recognition rate of the proposed method was at least 65.0%, reaching up to 96.7%. These results validate our proposed method.