Implementation of CNNs for Crop Diseases Classification: A Comparison of Pre-trained Model and Training from Scratch

In recent times, the machine learning field has become very progressive and has given impressive results, through the development of advanced Convolutional Neural Networks (CNN). These models simulate the biological behavior of the human-beings, especially, in dealing with the image recognition tasks. To train a deep CNN from scratch, a massive amount of labeled training samples are required that make the training process difficult. Also, the results have not converged properly and become overfitted in nature. Data augmentation is applied to overcome this issue of overfitting. An optimistic substitution is to use the pre-trained networks, prepared previously over a large dataset such as ImageNet. This kind of CNNs can also be deployed using their two alternative approaches, namely, feature extraction and fine-tuning of the pre-trained model. In this paper, 1296 leaf images of bean crops were used to perform the experiment that classifies the diseased or healthy leaf. In order to show the variation in the performance of the CNNs, the network has trained from scratch as well as using pre-trained networks. Experimental results showed that training from scratch performs worst (70% accuracy) over small training data and pre-trained networks (97.06%) gave far better results compared to the previous one. It is observed that the use of pre-trained networks with the tuning of hyperparameters is an optimal choice for training, for small training data set.