A facial expression recognition method based on the fuzzy information fusion of the local features is proposed. Each original image is divided into many sub-images and all training sub-images from the same position construct a new training subset. The traditional PCA (principal component analysis) operates directly on a set of new training subsets respectively and a set of projection sub-spaces can be obtained. The local sub-feature of an unknown face can be extracted by projecting each sub-image onto the corresponding sub-space. According to these local sub-features, the membership grades of the test sub-images to the training sub-images can be determined. The identity of an unknown facial expression image is determined by the fuzzy fusion which aggregates the local sub-features. The experiments on JAFFE database show the effectiveness of the proposed method.