USE_Res2Net:Automatic Plant Image Classification Using Squeeze and Excitation

Recent approaches and advances in computer vision have paved the way for employing deep and neural networks in wide-reaching image recovery tasks. One of the applications includes automated plant leaf disease classification. Crop diseases are a one of the important threat to the steady food supply for billions of people, and a substantial fragment of the crop fields are affected with the different type of diseases each and every year and thus are lost to them, leading to an immense loss in agricultural yield to millions of smallholder farmers around the globe. The main aim of this work is to predict or classify the most probable plant species (either with an affected disease or a healthy crop if no disease is diagnosed) for each and every observation of the images taken by the farmers with mobile. The proposed approach is named the USE_Res2Net model, which identifies different diseases. The input images collected from the PlantVillage dataset were few in number for some classes of crop diseases. In this proposed method, various image based data augmentation methods are used to avoid the over-fitting problem and improve the classification accuracy. The proposed model USE_Res2Net achieves an accuracy of 99% on standard dataset.