An improved fuzzy perceptron algorithm is suggested for the two-class classification. The basic idea originates from an observation that the variances of different moment invariants （features） under rotation

shift and size transform are not the same. Then the sample feature vectors are weighted

so that those features with small variance will play the dominant part. The algorithm is also extended to multiclass case and the experimental results are given. The computer simulation shows that the new method is more effective for the classification than the old one.