A Survey on Training Issues in Chili Leaf Diseases Identification Using Deep Learning Techniques
Kantha Raju Kanaparthi, S. Sudhakar Ilango
Abstract
The agricultural sector plays a crucial role in the majority of developing countries like India. But in recent times agriculture production is following a downward trend due to various plant diseases with an increase in investment costs. This work conducts a survey on deep learning techniques training issues related to a Chili leaf diseases dataset. Especially the work focused on the viability of the Squeeze-Net training architecture on the Chili leaves to train the two classes of diseased leaves namely Geminivirus and Mosaic. The dataset comprised of 160 Chili diseased photographs deployed from the Kaggle public domain is subjected to the Squeeze-Net convolutional neural network (CNN) to test the training accuracy. The obtained training accuracy ranges from 50% to 100% by considering various training properties like CNN optimizers SGDM, ADAM, and RMSPROP w.r.t Max_Epoches and assigning Dropout probability, Strides, Dilation factor, and padding values as constants. From the simulation is observed that the Squeeze-Net CNN architecture is achieving 100% accuracy in ADAM, and RMSPROP, where Max_Epoches are 40 and 35 respectively. But it suggested that the applicability of RMSPROP is good for training the Chili Dataset, where Max_Epoches are very less compared to the ADAM and SGDM.