Performance analysis of segmentation models to detect leaf diseases in tomato plant

Abstract

In agriculture around 22% of crop yield loss is due to living and non-living organisms such as biotic and abiotic stress/disease. The early-stage diagnosis of these stresses is an important issue for farmers through naked eyes. Using computer vision technologies can detect the pattern and clustering of diseases at an early stage. However, in recent times, deep learning technologies based on computer vision is helpful for the diagnosis of biotic stress (single biotic and multi biotic) in tomato plant leaves. In this work, the PlantVil lage dataset is gathered for the segmentation of object detection. The labeled, enhanced and augmented data has been used for training the model. The proposed hybrid Deep Seg mentation Convolutional Neural Network (Hybrid-DSCNN) model has been segmenting the diseased objects in the tomato plant. This Hybrid-DSCNN is assembled using U-Net and Seg-Net pre-trained models with instance segmentation for better detection of objects. The semantic segmented data has been recognized for the single and multiple leaf diseases for identifcation and classifcation in this work. A comparison of the predicted Hybrid DSCNN model’s output has been made with other modifed U-Net, M-SegNet, and modi fed U-SegNet in terms of Accuracy, Precision, Recall, and Intersection over Union (IoU), and mean Intersection over Union (mIoU). The proposed model processed 1004 images in 30 ns,which is better than other compared models. The accuracy achieved using the pro posed model is 98.24%, which is far better than other modifed segmentation models.