For the last few decades, the detection of diseases using deep learning architecture has gained huge popularity specifically for diabetic retinopathy. This disease progresses at different levels so early detection is crucial. The study proposes feature-based optimized deep learning neural network architecture for the detection of diabetic retinopathy. The features include a canny edge detector and histogram of oriented gradients. The proposed approach achieves an accuracy of 97.01% and 97.88% on publicly available, standard Messidor-2 and EyePACS datasets of Diabetic Retinopathy and outperforms existing state-of-the-art deep learning image classification Residual Network architectures of ResNet-50 and Inception-v3.