Diabetic retinopathy (DR) is among the most prevalent eye diseases that can result in blindness and vision loss if left untreated. Early detection of vision impairment might slow or stop its progression. Manual diagnosis using retinal fundus images, such as visual acuity testing, pupil dilation, and optical consistency tomography, calls for highly skilled clinicians to identify and assess the significance of numerous smallest details, making this a rigorous, time-consuming, and error-prone task. Consequently, a computer-aided automated process is definitely required. This study proposes an automated strategy for binary classification of DR versus normal retina using gaussian filtered color fundus retinal photos as input. The study employs the Diabetic Retinopathy dataset from Kaggle that includes 3,662 original retinal images with labelling for non-DR and DR. Using Convolutional Neural Network (CNN) architectures, the process can distinguish exudates, microaneurysms, and hemorrhages in retinal imaging by training with labelled data. The study trained and tested five models, InceptionV3, MobileNetV2, ResNet50, VGG16, and VGG19. It is observed that MobileNetV2 stands out as the most effective at detecting DR and Non-DR with and accuracy of 96.04% and Cohen Kappa Score of 92.08%. In the study, the models Resnet50, VGG16 and VGG19 have the same compilation time of 62 seconds for each epoch compared to which MobileNetv2 takes more time with 90 seconds, followed by InceptionV3 with 85 seconds.