Keratitis is one of the most common causes of permanent blindness in the world. Most blindness caused by keratitis can be avoided through early detection and treatment. In order to improve the current detection ability and treatment level of keratitis diseases, an automatic diagnosis method named dimensionality reduction on patch-based features (PBFs-DR) is proposed. Firstly, the target detection algorithm Faster R-CNN is used to locate the corneal region and conjunctival region of the patient's eye image; Then four patches were sampled on the obtained corneal region and conjunctival region respectively, and the high-dimensional features of each patch were extracted through Densenetl21 network; Finally, the features of all patches in the image are spliced, and the random forest algorithm is used for feature classification after dimension reduction. This method finally realizes the automatic diagnosis of keratitis, normal and other corneal abnormalities. Compared with the DenseNetl21 model based on traditional image-level, the accuracy is improved by 2.33%, and the automatic diagnosis level of keratitis is improved.