Improving the Performance of Convolutional Neural Networks (CNN) in Identification of Agricultural Plant Diseases
TL;DR
The findings confirm that the Inception V3 architecture provides a robust and efficient solution for automating plant disease diagnosis in the field and demonstrates that computational efficiency can go hand in hand with accuracy improvements on complex agricultural image datasets.
Abstract
This study aims to improve the performance of the Convolutional Neural Network (CNN) algorithm in detecting coffee leaf diseases using the Inception V3 architecture. The main challenge of this classification is the subtle visual similarity in color, texture, and symptom patterns between diseases. To overcome this, Inception V3 is implemented because of its superiority in multi-scale feature extraction through convolution factorization which reduces parameters while increasing accuracy. The dataset used consists of 1,120 images, evenly distributed into four classes (three types of diseases and one healthy class, each with 280 images), with a training, validation, and test data split ratio of 896:112:112. As a comparison, a conventional basic CNN architecture consisting of 3 convolution layers (3 X 3, stride 1), 3 max-pooling, and 1 dense layer, trained with the same hyperparameters (Adam optimizer, learning rate 0.001, batch size 32) is used. The experimental results show a significant performance improvement; Model accuracy increased from 74.4% on a standard CNN to 97.0% after integrating Inception V3. The scientific contribution of this research lies in mapping overlapping visual characteristics of coffee diseases through multi-scale feature optimization, which demonstrates that computational efficiency can go hand in hand with accuracy improvements on complex agricultural image datasets. These findings confirm that the Inception V3 architecture provides a robust and efficient solution for automating plant disease diagnosis in the field.