Abstract:  Tibetan character set is composed of 30 Tibetan letters, 4 Tibetan vowel signs, 4 Tibetan subscripts, 3 Tibetan superscripts，10 Tibetan digits and some Tibetan punctuation markers. All Tibetan regards syllable as the
word-building unit. Each syllable is separated by the syllable nodes and every horizontal unit of the syllable spelling is called a Tibetan character, It is the least one that the Tibetan character of a syllable counts, there can be 4 at
most.Tibetan character recognition takes Tibetan character as the basic recognition unit, so the difficulty of the printed Tibetan character recognition is how to select and extract appropriate features to represent a Tibetan
character. At present，the printed Tibetan character feature extraction mainly has two methods, namely image projection approach and directional line clement approach.The image projection approach obtains the character
characteristic vector from the character image pixels along a certain direction (such as vertical，horizontal or diagonal direction, etc.)projection.This method is characterized by that the algorithms of matching and classification arc
simple, easy to realize，and the anti-interference ability is strong, but its ability to distinguish similitude characters is bad.The extraction method of directional line clement feature is the edge of the character pixel by four directions;
horizontal 0º，vertical 90º，inclined 45ºand anti-oblique 135º to quantify and the quantization result is regarded as the direction attribute of the point. This approach not only includes the structure information of characters, but also has
the statistical property and its recognition performance is better than the image projection approach. But its feature vector dimension is too much, so use this method needs to take some compression algorithms，which makes this
method described comparatively complicated while identifying, the matching process complexity much higher. Approaches of feature extraction for printed Tibetan character both based on fractal moments and hased on the
improved rough grid are presented in this paper. After analyzing present feature extraction approaches一一image projection approach and directional line clement approach，the fractal moments theory and the rough grid arc applied
in the proposed approach.The extracted features using the proposed approach method can not only describe the local and global properties of the character, but also decrease the influence that caused by the decline in the rate of
recognition part due to the change of pixel’s position in the image.The methods can not only overcome high probability of misrecognition due to numerous Tbetan characters，but also solve the low run speed owing to plenty feature vector dimensions.