Res Net Driven Deep Learning Approach for Prediction of Potato Leaf Diseases

Abstract

Potatoes are a staple food, but their yields are greatly reduced by leaf diseases. Traditional approaches depend on experts who manually examine the plants, which is labor-intensive and expensive, thus they are not applicable for large-scale farming. To overcome this difficulty, in this paper, we propose a simple and fully automatic disease recognition scheme based on deep learning and image processing. Firstly, a quantity of potato leaves is take images. Following pre-processing, the images are cleaned, down sampled, and normalized to pixel intensities so that they can be used for training. The sample size is enhanced by applying data augmentation, such as rotation, flipping, zooming out and cropping (ZOC), and brightness modification, to enhance the model's robustness to different lighting conditions. The transfer learning technique is then applied for training to ResNet architectures: ResNet-18, ResNet-34 and ResNet-50. These have skip connections which allow them to learn superior features and prevent training problems at great depths. Each leaf is classified as healthy, early blight, or late blight. We assess the accuracy, precision, recall, and F1-score of the system. ResNet-50 obtains the best accuracy, ResNet-34 a good accuracy-speed trade-off and ResNet-18 is the fastest model and it is suitable for running on mobile devices. As the shallowest model of the Residual Networks, ResNet-18 contains the least number of parameters. In general, the proposed approach delivers a fast and accurate on-line solution for disease detection, so that a farmer can closely monitor and tend the crops and consequently raise the agricultural production.