Evaluation of Deep learning based Resnet-50 for Plant Disease Classification with Stability Analysis
Raj Kumar, Dinesh Singh, Anuradha Chug, Amit Prakash Singh
Abstract
Deep learning and machine learning techniques are extensively in use for plant disease recognition and their classification. The emergence of convolution neural networks had changed the earlier paradigm of image classification where separate image processing techniques were used for feature extraction. CNN-based feature extractors use convolution operation to automatically extract features from the input image. CNN architecture is inspired by a deep neural network which may have multiple hidden layers followed by a few dense or fully connected layers. The performances of deep neural networks are substantially affected by the training process as well as the amount and quality of training data being used. This research study evaluates the effects of hyperparameters and the training approaches on model performance. Here, we have assessed the exhibitions of four deep learning models (three models with training from scratch and one is a pre-trained model, i.e., ResNet50) on potato plant diseases early blight and late blight from the plant village dataset. As per the experimental results shown, models trained from scratch outperform (96.75 % and 94.43 % accuracy) as compared to ResNet50, pre-trained model (93.5 % accuracy). So, it has been observed that pre-trained models are not always a good choice for all datasets under study.