Overfitting Mitigation Analysis in Deep Learning Models for Plant Leaf Disease Recognition

Over the last two decades, deep learning models have revolutionized the field of image processing in almost every application area including the field of agriculture. Recently, these models have successfully been applied to recognize leaf diseases of plants. However, deep learning models are complex in nature and are prone to overfitting. Therefore, the selection of proper training settings is very important to avoid overfitting. In this work, we explore the effect of data augmentation and various optimization techniques being used to partially train a pretrained DenseNet-121 model from the perspective of overfitting. In first phase, experiments are conducted for different combinations of data augmentation techniques, it was found that for some combinations, the network is more prone to overfitting. In the second phase, seven different optimizers were analyzed and it was observed that the Adadelta optimizer exhibits the best generalization ability on the available training data.