Diabetic Retinopathy (DR) is a leading debilitate persistent malady and the main causes of blindness and visual wreckage in developed countries for diabetic victims. The research noted that 90% of the cases could be stopped through initial detection and cure. Initial eye screening through retinal images is used by doctors to detect the lesions related to DR. As the rising number of DR people, the amount of replica to be human evaluated is becoming stiff. Training new staff for this type of diagnosis is a very long process, because it requires better ability by daily practice.Diabetes arises when the beta cells (a type of islet cell) of the pancreas break down to secrete ample insulin, slowly altering the human eye’s retina. As it keeps going, the patient’s vision starts fading, noted to DR. It is most common in patients who’ve had diabetes for more than ten years.Researchers in the past proposed frameworks to classify retinal images into various categories of DR with limited success. In this, we try to conquer the hassle by using a hybrid deep learning model, which determines the pattern and divides the retinal images into one of the five grades.This methodology is applied to the APTOS 2019 Blindness Detection dataset of retina images taken using fundus photography under a range of imaging conditions.In this thesis, we studied various deep learning models for diabetic retinopathy level classification using retina images. For this purpose, multiple models based on automatic feature extraction and classification, based on deep neural networks were explored. In this thesis, two approaches are proposed. One approach is designed purely designed based on Capsule Network,while the other framework is based on a combination of Convolutational Neural Network and a Capsule network. In this paper, we have proposed DRDNet (Diabetic Retinopathy Diagnosis Network), a neural network framework based on capsule networks (CapsNets) for DR diagnosis.Experiments on a dataset with 1,265 images demonstrate that CapsNet shows better accuracy and convergence behavior for the complex data than the state-of-the-art techniques. The proposed DRNet performs with an overall accuracy of 80.59% for five classes compared to the closest competitor with an accuracy of 75.83%. We conducted a study on a mixed dataset for two types and found that testing accuracy was 80.59%. We have also done training on a two-class model and testing on other unseen datasets. Moreover, we observed that DRDNet has much higher confidence for the predicted probabilities than other state-of-the-art techniques.The second model we named as DRISTI (Diabetic Retinopathy classIfication by analySing reTinal Images), where a hybrid deep learning model composed of VGG16 and Capsule network is proposed. DRISTI determines the pattern and divides the retinal images into one of the five grades.This methodology is applied to retinal images from publicly available datasets. The results, so obtained, are a 82.06% validation accuracy and a 75.81% testing accuracy for 5 class classification and 96.24% validation accuracy and a 95.50% testing accuracy for 2 class classification using VGG16.We have also performed cross and mixed dataset testing to demonstrate the efficiency of DRISTI.In this thesis, we have proposed such a model for the first time, which has given us better results than all the models we have compared with. To validate our claim we have reported through experimental study.In this thesis, we also make a web-based diabetic retinopathy classifier that predicts that if someone has symptoms of DR or not within a second, you have just uploaded the retina image of the eye and clicks the predict button for the result. This web-based classifier is best for the initial screening of diabetic retinopathy