Diabetic retinopathy (DR) is a crucial complication responsible for visual impairment and even blindness in the patients with diabetes. The most important way to prevent severe eye-related outcomes is early detection and classification of DR stages. In this research, we present a new image fusion oriented technique for identifying and categorizing different levels of DR with the information from both eyes. The process adopts a novel fusion algorithm Gradient, Spatial frequency, Entropy, and Standard deviation (GSED) which calculates the image weight function for fusion using 4 parameters namely, Mean Gradient, Spatial Frequency, Entropy, and Standard deviation. The left and right eye images are combined using Weighted Average scheme generating final fused image. The proposed method was evaluated with publicly available datasets of retinal images against state-of-the-art automated DR grading approaches. EfficientNet produced the highest accuracy of 83.55%, which is the best performing model in terms of percentage achieved on the test data set.