Data from: PMDS-YOLO: A Lightweight Deep Learning Model for Non-Destructive Citrus Fruit Diameter Estimation and Quality Grading

The original dataset includes 1287 raw images. After data augmentation (random horizontal flip, scaling, rotation, random occlusion, brightness‑contrast‑saturation adjustment), the total dataset was expanded to 6435 images. All augmented images have synchronously updated YOLO‑format bounding‑box annotations.

Annotation was performed using LabelImg software with rectangular bounding boxes. Two object classes are defined: ‑ Orange: citrus fruit instances ‑ Coin: 25 mm‑diameter standard reference coin For partially occluded fruits, annotations were drawn only according to the visible fruit region. Annotations are saved as YOLO‑format .txt files with file names matching corresponding image filenames.

Dataset split follows a 7:2:1 ratio for training / validation / test subsets. All identifiable images belonging to one physical fruit are assigned to one single subset to prevent sample leakage across subsets and guarantee sample independence. Annotation statistics (after augmentation): ‑ Training set: 15317 Orange, 2872 Coin ‑ Validation set: 4405 Orange, 822 Coin ‑ Test set: 2582 Orange, 380 Coin Ground‑truth measurements for 100 citrus fruits are provided, including manually measured transverse diameter (mm) and single‑fruit weight (g), acquired by vernier caliper and electronic balance, for regression‑based diameter calibration.

Missing values were filled by mean imputation; outlier samples were directly removed. File structure: images/train, images/val, images/test labels/train, labels/val, labels/test measurement_100_samples.csv License: CC BY 4.0 Associated publication: PMDS‑YOLO: A Lightweight Deep Learning Model for Non‑Destructive Citrus Fruit Diameter Estimation and Quality Grading, Foods.