This study classifies the stage of Diabetic Retinopathy (DR) into three classes, namely normal, mild Non-Proliferative Diabetic Retinopathy (NPDR), and moderate/severe NPDR class. In general, this research is done to solve that problem arises as a result of similarity of image per stages that cannot be assessed invisible. So, it requires a handling where the image of the retina can be categorized into appropriate categories. Based on the problem, two experimental mechanisms were conducted for each hierarchy, i.e approach computer vision that only focus to process the whole image and the approach taken by the medical using texture feature as marker feature to detect DR. The data are obtained from DiaretDB0 public database. In this research, we present Histogram of Oriented Gradients (HOG) to extract feature. To select the best feature from HOG, we used factor analysis as a feature selection method. This step was done to get good performance in classification step. In our experimental design, we implented shallow learning such as Support Vector Machines learning and Random Forest learning to classify moderate/sever NPDR vs. mild NPDR, mild NPDR vs. Normal, and moderate/severe NPDR vs. Normal. The experimental result shows that our proposed method is able to provide good enough performance in terms of time and accuracy. Our proposed method achieved around 85% accuracy for the binary class classification.