Attention-Augmented Multi-Scale CNNs for Robust Plant Leaf Disease Detection and Classification

Abstract

The growing influence of plant diseases on agricultural productivity becomes a major threat to the food security of the world. Traditional methods of identifying plant diseases of leaves are time consuming, prone to mistakes and overwhelmingly dependent on the knowledge of experts to make a decision. In this project, an Attention-Augmented Multi-Scale Convolutional Neural Network (CNN) deep learning-based method for automated detection of plant diseases is proposed. The model combines multi-scale feature extraction with channel and spatial attention, which can better learn discriminative feature representations from leaf images. Multi-scale convolutions are adopted which learn both the small-scale and large-scale disease patterns, and the Squeeze-and-Excitation (SENet) and the Convolutional Block Attention Module (CBAM) are taken into account which make the network only pay attention to the leaf regions via the veins, edges, and infected spots, and decrease the background noise interference. Images are preprocessed with the background removal and augmentation techniques to improve generalization. A notable fact is that performance of the proposed architecture outperforms the traditional CNN models in terms of significant enhancements in classification accuracy and classification robustness. The implementation for the system is TensorFlow & Keras AI framework using GPU accelerated Colab environments for fast convergence of the system. Using the help of a Streamlit-based web app for real-time prediction of diseases, users can upload leaf images and can get the instant diagnosis. The presented end-to-end system shows how attention-augmented networks play a major role in aiding early disease detection in precision agriculture for sustainable crop management and mitigation of economic losses.

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.