Region Proposal Convolutional Neural Network with augmentation to identifying Cassava leaf disease

Abstract

This article describes a new idea in recognizing cassava plant disease patterns based on the damage that occurs to the leaves. Classification using an image processing approach is used to solve these problems. The aim is to improve the classification results that have been carried out by previous researchers. There are four classes of observed disease and one class of Normal. Based on the image resources of cassava leaves, sometimes there are background colors that are almost the same or close to the color of the object being sought, so a solution with the right region contour method is needed. The proposed region uses the Convex Hull approach. The results showed that better accuracy values were obtained by using a Convolutional Neural Network with a region. The addition of the proposed region clarifies the area observed in cassava leaves. The proposed Convolutional neural network method can recognize patterns well in the previous architecture and also in the Custom Layer. The addition of the regional proposed method increases the classification accuracy indicator by an average of 99.01%. Evaluation of the effectiveness of this method was confirmed by calculating the average MSE 0.0080, RMSE 0.0935, and MAE 0.0063 with an average training computation time of about 7 minutes 50 seconds.