Abstract It is a important issue how to separate text from scanned image of engineering drawings. The paper presents an algorithm to extract characters and their features from images using Primitive Region Adjacency Graph. We can easily get character stroke regions from primitive regions. At start of a character stroke region, a character region grows by traversing the graph for adjacent character stroke regions. After analyzing features of character regions, we can get real character regions. A string can be got through combining those near and collinear character regions. If rectangles inflated of two character regions are intersected, they are near. The direction of a string region is attained by the center points of enclosed rectangles of characters shared by the string and figures attached. Then characters not horizontal are rotated to horizontal. A Primitive Region Adjacency Graph can represent geometrical and topological features of a character region, which is helpful to extract features of a character region. Some applications show that the algorithm can deal with adherence of characters to graphics, and is effective and robust.

Key words Primitive Region Adjacency Graph, text-graphics separation, character stroke regions, character extraction, feature extraction