Diabetic retinopathy (DR) is a type of diabetes mellitus that attacks the retina of the eye. DR will cause patients to experience blindness slowly. Generally, DR can be detected by using a special instrument called an ophthalmoscope to view the inside of the eyeball. However, in conditions where there is a very small difference between the normal image and the DR image, computer-based assistance is needed for maximizing image reading value. In this research, a method of image quality improvement will be carried out which will then be integrated with a classification algorithm based on deep learning. The results of image improvement using Contrast Limited Adaptive Histogram Equalization (CLAHE) shows that the average accuracy of the method on several models is very competitive, 91% for the VGG16 model, 95% for InceptionV3, and 97% for EfficientNet compared to the results original image which only has an accuracy of 87% for VGG16 model, 90% for InceptionV3 model, and 95% for EfficientNet. However, in ResNet34 better accuracy is obtained in the original image with an accuracy of 95% while in the CLAHE image the accuracy value is only 84%. The results of this comprehensive evaluation and recommendation of famous backbone networks can be useful in the computer-aided diagnosis of diabetic retinopathy.