Orange Fruit Image Dataset for Classification

This dataset was collected to enable the development and benchmarking of lightweight deep learning models for on-device orange quality assessment. A total of 15,000 labeled images were acquired using standard smartphone cameras in Bangladesh, covering five agronomically relevant classes: fresh, black‑spotted, canker‑affected, green, and rotten oranges. Each class contains exactly 3,000 images, ensuring a balanced dataset that avoids classifier bias toward majority classes.

Images were captured under variable real‑world conditions (different lighting, backgrounds, angles) to simulate the challenges of field deployment. All images were resized to 224×224 pixels and stored as RGB JPEGs. To further enhance model robustness, the authors applied an automated augmentation pipeline (rotation, shifting, zoom, shear, brightness, and channel adjustments) during training. The dataset was split into training (70%), validation (15%), and test (15%) sets using stratified sampling for reproducibility.

The dataset was used to train six models: a custom EinsteinNet, ResNet50, DenseNet121, MobileNetV2, NASNetMobile, and a Google Teachable Machine baseline. The best offline test accuracy reached 99.87% (Teachable Machine) and 99.6% (EinsteinNet). On‑device testing on a Google Pixel 6 confirmed that the dataset’s variability supports real‑world accuracy >95% for most models, though trade‑offs between latency, model size, and power consumption were observed.