Fast Convergence of Lightweight Deep Learning Models for Plant Leaf Disease Recognition

Disease recognition in plants have been a key focus of researchers around the globe to improve the production quality of vegetable, fruits and crops. Deep Learning techniques are implemented to classify biotic and abiotic plant diseases. Numerous studies focused on accurately identifying the visual features and in the quest to improve the accuracy, the networks get deeper and deeper and hence requiring more computational resources, time and effort. The main focus of our proposed work is to reduce the computational complexity and cost for a deep network proposed for automatic plant leaf disease recognition with the help of lightweight convolutional neural networks (CNNs). Since computational cost of the CNNs directly depends on the number of convolutional operations performed, we in this paper present a hyper-parameters based tuning strategy to improve three parameters namely validation loss, accuracy and speed of convergence. Experimental results show promising results for various pretrained models available for experimentation in Keras on online Kaggle platform.