The Convolutional Neural Network for Plant Disease Detection Using Hierarchical Mixed Pooling Technique with Smoothing to Sharpening Approach

Abstract

Plant health is an important factor in agricultural production as it mostly affected by plant diseases. Due to plant diseases, the growth and crop yield gets affected which results in negative impact on agriculture in terms of economic loss to farmers. In plant disease management, early and accurate disease detection can control its spreading and avoid unnecessary loss to farmers. Traditionally, plant disease detection has been carried out through visual inspection by human experts. This method is based on subjective perception hence it has risk for error in detecting accurate disease. In recent past, researchers have proposes numerous machine learning approaches to detect the plant diseases. Due to advancement in artificial intelligence and electronic gadgets technology, there is large scope for improvement in neural network algorithms for detecting plant diseases early and accurately by extracting leaves features efficiently. To detect tomato plant diseases, the novel convolutional neural network (CNN) model has been proposed in this paper. The hierarchical mixed pooling technique for smoothing to sharpening approach has been used in proposed CNN model. The system uses tomato plant leaf images obtained from Kaggle dataset. The system has been trained with 1000 images of healthy leaf and 1000 images each for nine different diseases frequently occurs in tomato plant. The different training models has been framed and experimented to identify efficient hierarchy of pooling techniques. The CNN training model_3 exhibit smoothing to sharpening approach with "Average-Max-GlobalMax" mixed pooling hierarchy and depicts better performance with a training loss 28.88%, a validation loss 12.61%, a training accuracy 96.46%, and a validation accuracy 95.41% at 20 epochs. Also, the performance of designed system have been evaluated with different state-of-art deep learning algorithms and compared with proposed CNN model.