Diabetic retinopathy is a disease in which the retina can damage completely by rupturing, bleeding blood vessels and a person can tend to suffer blindness if not cured in early stage. Diabetic retinopathy disease can be cured by detecting the disease at a mild stage and taking proper treatment by medical professionals. In this work, we have classified four stages of diabetic retinopathy using deep learning. Deep learning is helpful to develop a model which can classify the image data with good performance, it works with pixel values of the image and each layer in the neural network detects the data from the images. In this work, we processed image data with preprocessing and contrast enhancement and histogram equalization (CLAHE Method) to equalize the contrast in the images. We extracted features using various methods like a firefly, curvelet, and scale-invariant feature transform methods. These images are feed to a convolution neural network and multilayer perceptron for the classification of diabetic retinopathy into mild, moderate, severe, proliferative stages. Max pooling layer in convolutional neural network extracts only low-level features from the image data. Therefore to establish the relationship between low-level features and high-level features we used different feature extraction methods.By modifying the convolutional neural network max pooling layer with feature extraction methods. From the results, we can conclude the best feature extraction method with a neural network to detect and classify diabetic retinopathy.