Trachoma is the leading bacterial infectious cause of blindness worldwide especially in developing countries such as Ethiopia. Examination for clinical signs of trachoma involves careful inspection of the lashes, cornea, eversion of the upper lid, and the tarsal conjunctiva. Ophthalmologists uses a slit lamp camera (if available) to inspect the clinical signs. Various researches have been done to automate detection of eye diseases. Almost all studies were concentrated on detection of diabetic retinopathy, glaucoma, and in a lesser extent to cataract. Automatic detection and grading of trachoma was neglected in previous studies. Thus, in this work, we present a system for automatic detection and grading of trachoma in line with World Health Organization (WHO) guideline. The proposed system has four components: preprocessing, segmentation, feature extraction and classification. In image preprocessing, we normalize the image to a standard size. Segmentation is used for region of interest (ROI) extraction and we propose a novel segmentation algorithm to extract the ROI (oval shape) of the eye. We use predetermined coordinate obtained empirically from the dataset. In feature extraction, we propose to apply Gabor filter on the raw image for texture feature extraction. It is used for detecting and selecting important features that account for the symptom of the disease. For classification we use convolutional neural network. A 4-way Softmax is used for grading into a specific class (normal, trachomatous scarring (TS), trachomatous trichiasis (TT), and corneal opacity (CO)). The proposed system is implemented using Keras (using TensorFlow as a backend) in Python and tested using sample image dataset collected from Carter Center Ethiopia. The model achieved a diagnosis accuracy of 98% for training and 97.9% for testing to detect and grade trachoma. Our model was faster to train and had smaller model size as compared to the state-of-the-art models. In addition, the application of Gabor filter and segmentation are also used to improve the performance of state-of-the-art models as well by 3% (AlexNet) and 1% (GoogLeNet).