An Optimized Plant Disease Detection Convolutional Neural Network for Future Hardware Implementation

Abstract

The potato is considered one of the most important crops, especially in Egypt. It is grown on mostly all continents, with a noticeable increase in production and consumption in Africa and Asia. This crop is facing a lot of challenges that lead to a reduction in the quality and quantity of the yield. One of the biggest challenges that affects the potato crop is the plant’s diseases. Potato early blight disease can cause severe damage to both potato foliage and tubers, and its damage can range from 5% to 50% loss in yield. Another important and historically significant potato disease is late blight. It is believed that the loss from such disease may reach up to 16% of the total global potato production. Due to the significant rise of deep learning, convolutional neural networks have been intensively used in tasks related to classification and image detection, with remarkable outcomes. Consequently, this work aims to implement a 2-convolutional layer neural network that could be implemented on a hardware platform, especially FPGAs, for its re-programmability and its ability to reduce overhead functions in CPU and GPU systems. The proposed model in this work with only 2 convolutional hidden layers proves to be more accurate than VGG16 and Resnet50 with a testing accuracy of 95.32%, in addition to its minimum resources, which in future work will make it more suitable to be implemented on FPGAs.