The diabetes-related eye disease called diabetic retinopathy (DR) affects the retina, the light-sensitive tissue located in the back of the eye. It is a leading cause of blindness in adults and a common complication of diabetes. Therefore, early identification of DR is crucial, especially in its non-proliferative stages. Fortunately, recent technological advancements, specifically in artificial intelligence (AI) and medical imaging, have led to the development of several methodologies for the early diagnosis and management of DR. However, the majority of approaches ignored segmentation and feature reduction stages, which are recognized as a significant research gap in the DR detection. Hence, in this research paper, at first, we preprocess the retinal fundus images using histogram equalization (HE), and then we remove the optic disc or optic nerve head by circular Hough transform. Subsequently, gray-level thresholding followed by top hat transformation, Hough transform, and Gabor filtering (GF) techniques are employed to eradicate blood vessels and find the abnormality of fundus images. Later, we generate a feature vector by extracting the gray-level co-occurrence matrix (GLCM) and scale-invariant feature transform (SIFT) features. This feature vector passes through the Modified Rider Optimization Algorithm (MROA) to generate optimal features. Finally, the extracted features are fed to a Recurrent Neural Network (RNN) for categorizing them into normal, early, moderate, and severe stages. This Research paper contributes to Improved diagnostic accuracy, reducing false positives and negatives and utilization of advanced imaging techniques that provide more detailed or earlier detection of retinal images. The experimental results revealed an accuracy of 92.3%, which is higher than the well-received strategies.