A Handheld Device for Apple Ripeness and Sweetness Grading with Convolutional Neural Network

Abstract

This research describes a handheld system developed for capturing multispectral images of apple fruit to evaluate quality based on ripeness and sweetness. The device consists of a 7-inch touchscreen display powered by a Raspberry Pi 5 and a rechargeable power bank, offering flexibility for field operation. A USB digital microscope camera with 1000× magnification is used to acquire detailed images. Multispectral data is captured using a Digitek lighting device, which provides sequential illumination in six color wavelengths: red, yellow, green, cyan, blue, and magenta. Image capture is automated through Python-based code that controls the lighting sequence and image acquisition. To grade apples by ripeness, a DenseNet-121 convolutional neural network was trained on the collected images, achieving an accuracy of 73.77%. For grading by sweetness, a multi-architecture approach, evaluating both custom convolutional neural networks (CNNs) and transfer learning models based on pre-trained VGG16, ResNet50, and EfficientNetB0 architectures fine-tuned on ImageNet weights was employed, reaching an accuracy of 66.22%. The results demonstrate the feasibility of using low-cost, handheld device paired with deep learning techniques for non-destructive fruit quality evaluation. The system holds promise for on-site grading in agricultural environments, reducing the dependence on hard-to-use and costly laboratory equipment.