Long term diabetes could cause the diabetic retinopathy (DR) disease, if in long term is not treated appropriately, could affect in loss of vision and the effect is irreversible. Automated DR grading is important to help the ophthalmologist in DR treatment. Exudates detection is part of DR detection, but unfortunately, the exudates detection mostly focusses on binary grading (disease or no disease) rather than multi-class grading. Thus, in this paper, we proposed method in multiclass DR detection by using exudates candidates. Exudates candidates can be obtained by utilized CLAHE and wiener filter to enhanced the fundus images. Then to improve the candidates, region growing, segmentation and clustering methods which considering circularity, areas and eccentricity are utilized. Finally, Those candidates were extracted for features using statistical features and fed into ensemble learning process. The results concluded that our methods with XG-Boost as ensemble classifier is able to grade the multi-class DR severity level and comparable with other researcher which use MESSIDOR datasets