This paper presents an automatic diabetes Retinopathy (DR) detection system using fundus images. The proposed automatic DR screening model saves the time of the ophthalmologist in disease diagnosis. In this approach, the segmentation is conducted using an improved watershed algorithm and Gray Level Co-occurrence Matrix (GLCM) is used for feature extraction. An improved Ensemble Extreme Learning Machine (EELM) is used for classification and its weights are tuned using the Crystal Structure Algorithm (CRYSTAL) algorithm which also optimizes the loss function of the EELM classifier. The experiments are conducted using two datasets namely DRIVE and MESSIDOR by comparing the proposed approach against different state-of-art techniques such as Support Vector Machine, VGG19, Ensemble classifier, and Synergic Deep Learning model. When compared to existing methodologies, the proposed approach has sensitivity, specificity, and accuracy scores of 97%, 97.3%, and 98%, respectively.