Hybrid Activation Functions in Deep Convolutional Neural Networks for Maize and Paddy Leaf Disease Recognition: A Transfer Learning Approach

Early detection of crop diseases is an art to prevent further loss in the yield. It not only prevents further spread but also increases the revenue for the farmers. Leaf is a major part of any plant and its good health is very important for maximum yield. Maize and paddy are major food crops grown in India and contribute significant part in food security. There are various reasons for leaves to get diseases that include pathogens like virus, bacteria, pests and even adverse environmental conditions. The idea of proposed work is to take dataset of maize and paddy leaf images with diseases like corn blight, corn common rust, corn gray leaf spot, brown spot, hispa, leaf blast along with healthy leaves. Use transfer learning to train different Deep Convolutional Neural Networks (DCNNs) with hybrid activation functions and use the trained models to test new leaf images. Different activation functions like Rectified Linear Unit (ReLU), Exponential Linear Unit (ELU) and Leaky Rectified Linear Unit (Leaky-ReLU) have been combined and applied. Initially all DCNNs are trained by using respective original activation functions, later hybrid activation functions are applied with suitable epochs, learning rate, iterations. Different DCNNs like AlexNet, DarkNet-19, VGG-16, SqueezeNet, and ResNet-18 have been tested on maize leaf images and obtained accuracies from 75 to 97.5%. Further, Shuffle Net, DarkNet 53, and Resnet-50 have been used for paddy leaf images and obtained accuracies from 70 to 98%. Our study found that AlexNet and VGG-16 and ResNet-50 show promising improvements on classification accuracy with hybrid activation functions. A detailed ensemble of DCNNs is given at the end to check the performance of various hybrid activation functions.