Sistema de inspeção visual para detecção de sarna em folhas de macieiras

Abstract

Apple scab, caused by the fungus Venturia inaequalis, poses a substantial threat to apple production, particularly for small-scale farmers in resource-constrained environments. Traditional visual inspection methods are often ineffective, necessitate technical expertise, and are impractical on a large scale. This work proposes an economical and offline-capable solution for detecting apple scab, employing lightweight Convolutional Neural Networks (CNNs) and an intuitive web application. Images were collected from three datasets—ATLDSD, eScab, and AppleScabLDs—and class imbalance was mitigated through data augmentation. Various models were evaluated, including MobileNet V3, ShuffleNet V2, and RegNet, with ShuffleNet V2 achieving the highest accuracy (92.83%). The system integrates a web application based on Next.js and React.js, hosted on Vercel, with local inference utilizing ONNXRuntime-web and Jimp for image preprocessing. This solution provides farmers with an accessible, efficient, and offline diagnostic tool, contributing to enhanced agricultural productivity and improved disease management.