Recognizing characters in tables of low resolution is a difficult problem because characters are too small relative to scanner resolution and adjacent characters are connected together. Input image should be segmented into isolated single characters. In this process, character images are seriously distorted. Approach for recognition of this kind poor image is different from these of high resolution.
In this thesis, we describe a system for table recognition of low resolution. The system aims to recognize numerals as well as some special symbols in tables such as stock lists. First, this system segments a table image into semantically meaningful region such as title, explanation and table-unit. The system extracts items from table-unit considering rule lines. Using the information of boundary and height, a given item is separated into isolated character images. Numerals are recognized by matching stroke templates. After detecting existence of obvious strokes, candidate numerals are selected as hypothesis and required existence or non-existence of storkes is confirmed. On the other hand, symbols are recognized by checking boundary shape. In a test with stock tables in a newspaper, the system achieves 97\% of correct recognition rate of numerals and symbols.
