Performance Evaluation of Semantic Segmentation Models for Identification of Sweet Potato Leaf Diseases

Semantic segmentation models have been proposed to identify plant leaf diseases. However, these models need to be evaluated on different datasets and applications for validation of model performance and deployment on edge devices. This research carried out performance evaluation of U-Net, Compressed U-Net, SegNet and PSPNet segmentation models to identify early and late blight diseased leaf images of Sweet potato. The image dataset was obtained from Plant village dataset. Three hundred leaf images were manually annotated using VGG Image Annotator (VIA) tool. Python programming language version 3.10.6 was used to implement the models. The dataset was split into 70% for training, 20% for validation and 10% for testing. The results of evaluation of the models showed that the Compressed U-Net outperformed other models for the dataset. The compressed U-Net requires only 7.9% of space needed by standard U-Net model and inference time for predicting the potato diseased leaves is good without compromise with mean intersection over union. This work recommends evaluating the performance of Compressed U-Net on diseased leaves of different crops and deploying the model on edge devices such as Raspberry Pi, smartphones to validate the model results.