U2-Net Segmentation for Bean Leaf Disease Classification: A Statistically Validated Evaluation with DenseNet121

Abstract

Prior work by the authors proposed a modified DenseNet121 architecture for classifying bean leaf diseases and reported a test accuracy of 96.90% on the raw iBean dataset (1,295 images) using a single evaluation run. Because bean leaf images are often captured against complex, cluttered backgrounds, and because a single-run evaluation does not indicate how stable a model's performance is, this study extends that work along two directions that were not addressed previously: (1) an image segmentation stage, comparing U-Net and U2-Net, is introduced to isolate the leaf object from its background prior to classification, and (2) model performance is validated through 30 independent repeated runs and paired t-tests rather than a single evaluation. U2-Net achieved a substantially higher mean Intersection-over-Union (mIoU = 0.98) than U-Net (mIoU = 0.61) and was therefore used to produce a cleaned, class-balanced dataset of 1,233 images (411 per class). Using the same classification head configuration as the prior work (DenseNet121 backbone with two additional convolutional blocks), the model achieved a mean test accuracy of 94.63% (SD = 1.16) over 30 runs, which was significantly higher than standard DenseNet121 (92.53%, SD = 2.00, t(58) = 4.98, p < 0.0001), InceptionV3 (85.60%, SD = 3.12, t(58) = 14.84, p < 0.0001), ResNet152V2 (87.20%, SD = 2.38, t(58) = 15.36, p < 0.0001), and ResNet50V2 (86.70%, SD = 2.74, t(58) = 14.59, p < 0.0001). On its best single evaluation run, the model reached a test accuracy of 97-98%, with precision 0.97, recall 0.98, and F1-score 0.98. These results show that, on a segmented and class-balanced dataset, the previously proposed classifier maintains high and statistically stable performance across repeated trials. Because the dataset, hyperparameters, and evaluation protocol differ from the earlier study, this paper does not claim that segmentation alone caused the accuracy difference, instead, it reports the combined effect of these changes and identifies isolating the individual contribution of segmentation as a direction for future work.