Automatic Early Diagnosis of Dome Galls in Cordia Dichotoma G. Forst. Using Deep Transfer Learning

There are several infections and diseases that can affect plants. Diagnosis of plant diseases is a challenging task. A fascinating method for identifying plant diseases is computer-based diagnosis using digital images of plant’s leaves. Most of the earlier research in this field has been devoted to feature engineering and traditional machine learning (ML) techniques. Based on hand-crafted features taken from digital photographs of the plant’s leaves, these methods identify various plant diseases. It can be challenging to use feature engineering to extract high-quality features from these digital images. Deep learning (DL) algorithms, have relieved feature engineering by automatically extracting and learning resilient features. However, the high number of parameters in conventional DL models typically leads to overfitting. Therefore, gradient vanishing problems in vast networks intensify learning failure and generalization errors. Additionally, getting a big dataset for starting from scratch to train a deep learning model is also difficult. To solve these issues to detect Dome Galls in
Cordia dichotoma
G. Forst. early, this research suggests quick and efficient deep transfer learning (DTL) model. In the proposed method, a custom dataset of
Cordia dichotoma
leaf images is created. 1784 images of
Cordia dichotoma
leaf are collected in real world environment, and offline augmentation techniques are applied for obtaining the final training dataset of 5400 images. Further, image preprocessing techniques are used to enhance the images. A DL model, Yolov4, has been modified and is trained for the early detection of dome galls in the leaves of Cordia. The model is trained using pre-trained weights, a process called transfer learning (TL). The model is tested on another set of 200 images, and the results show an accuracy of 95% and an F1-score of 95.8%. Experimental results show that the modified Yolov4 performs 3% more accurately than the original Yolov4.