Semantic Segmentation Based Leaf Disease Severity Estimation Using Deep Learning Algorithms

With the advent of deep learning algorithms research work in object recognitions has produced high quality algorithms that outperforms classical image processing techniques. In this work we are proposing a novel approach which employs semantic segmentation to estimate the severity of the leaf disease. For semantic segmentation we have used a light weight deep learning architecture SegNet. Primarily the SegNet removes the background noise and in its subsequent phase it locates the necrotic scars/lesions caused due to leaf diseases and performs semantic segmentation. The estimation of amount of damage caused to the leaf depends on the diseased region/part of the leaf. Through SegNet the proposed work identifies the healthy region and diseased region of the leaf and pixel-level labeling is done. When compared SegNet with other deep learning based semantic segmentation architectures like FPN, Unet and DeepLabv3, SegNet proves to be memory efficient as it stores only the max-pooling indices of the feature-maps. Further this works extends the architecture for classification problem using ResNet. Moreover in the proposed work the accuracy levels of the disease severity obtained are very close to the manual methods and satisfactory.