Mechanical part recognition has great significance in automated sorting and processing. The core of mechanical part recognition is the part image feature extraction. Thus, how to extract part features to meet the real time requirements for the automated production line has a crucial role. A novel feature extraction algorithm is presented for part image features in this paper. It aims to optimize the fuzzy Fisher criterion function to figure out two orthogonal optimal discriminant vectors in an unsupervised way. Based on these two vectors, the linear transformation from d-dimension to 2-dimension can be obtained. Experimental results on three mechanical parts show its effectiveness.