This paper describes a structure analysis method of character regions on the backbone of a book extracted from a bookshelf image by the image processing method previously proposed by the authors. This analysis method consists of two processes. In one process, character string regions are made from extracted character regions by the use of size and position of circumscribed rectangles of the characters. In the other one, the character strings are classified into several categories such as title of a book, author's name and so on, by a 4-layered neural network whose input patterns are geometrical features of the character strings. Correct recognition ratio for 497 regions of 115 books was 89.5 percents in the region recognition experiments. This method can be applied to making library databases, automatic management of books and so on.