Transformer-augmented EfficientNetV2B3 for robust plant disease identification
TL;DR
A hybrid image classification deep-learning model using Convolutional Neural Network EfficientNetV2B3 combined with Transformer block made of Multi-head Attention and Multilayer Perceptron (Feedforward layers) generalizes better and performs well in detecting plant diseases.
Abstract
Plant-leaf diseases cause a significant damage to the agriculture yield if they are not diagnosed and treated early. It is also crucial in preserving global food security and promoting sustainable farming practices. These diseases can be detected through manual inspection but it is laborious to do by hand and the outcome is entirely dependent on the examiner. It has been observed that manual assessment prone to errors, particularly when there are irregularities in the illumination, abnormalities in the leaves, and small variations in disease symptoms. So, there is a need for a model that can successfully classify data by extracting features using computer vision and deep learning. This paper introduces a hybrid image classification deep-learning model using Convolutional Neural Network EfficientNetV2B3 combined with Transformer block made of Multi-head Attention and Multilayer Perceptron (Feedforward layers). EfficientNetV2B3 known for its scaling efficiency is used as a backbone for initial feature extraction, while the multi-head attention lets the model to learn relationships between distant regions by focusing on multiple areas of the image and the feedforward layers help model to figure complex features and then classified through a softmax output layer. The study tells that this model with less parameters, speed and high accuracy than existing image classification models like Resnet, VGG, Inception etc. generalizes better and performs well in detecting plant diseases with a validation accuracy of 99.70%.