Novel Convolutional Neural Network that Uses a Two-Stage Inception Module for Bacterial Blight and Brown Spot Identification in Rice plant

摘要

In this paper, a high accuracy and low computational budget Convolution Neural Network (CNN) architecture is proposed for detecting bacterial leaf blight and brown spot in rice plant. A combination of image preprocessing techniques followed by a low parameter count CN architecture is presented. This architecture gives considerably better accuracy on a limited dataset and makes the classifier learn only what is needed. It also uses significantly lesser computational resources than the present state-of-the-art. The mixed approach overcomes multiple issues like background noise, limited dataset availability and a large network requirement. The selection and development of the deep learning architecture and the pre-processing algorithms were the key aspects of this study. The architecture is fine-tuned to suit present mobile applications which not only seek high accuracy, but also low memory and computational costs. Data augmentation is used to enhance the size and quality of the limited dataset. As seen by the results of 3250 pre-processed and data augmented images, the train and test accuracy for detecting leaf blight and brown spot are 99.5% and 97.3% respectively. Total parameter count of 1.2 million and computational count of 448.49 million, indicate clearly, a significantly improved performance than the present state-of-the-art.