An annotated image dataset for small apple fruitlet detection in complex orchard environments.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Horticulture

Background:

• Automated systems for fruit production require high-quality datasets for accurate detection.
• Small fruit detection is crucial for pre-thinning processes to optimize yield and quality.
• Existing datasets may lack sufficient variation for robust small apple detection in diverse orchard conditions.

Purpose of the Study:

• To introduce a novel dataset specifically for small apple detection in pre-thinning applications.
• To provide a standardized resource for training and validating intelligent thinning systems.
• To facilitate advancements in automated fruit production within the apple industry.

Main Methods:

• Collection of 2,517 RGB images under varied real-world orchard conditions (weather, lighting, fruit size 3-25mm).
• Standardization of images to 500x500 pixels from original 3024x3024 resolution.
• Annotation of small apple targets using LabelImg software with PASCAL VOC (XML) and YOLO (TXT) bounding box formats.

Main Results:

• The dataset effectively supports the training and validation of various object detection architectures (e.g., Faster R-CNN, YOLO, RT-DETR).
• Experiments confirm the dataset's utility in demonstrating the performance of multiple detection models.
• The annotated data enables precise localization of small apples within the specified size range.

Conclusions:

• The introduced small apple pre-thinning dataset is a valuable resource for developing automated thinning systems.
• This dataset can significantly contribute to enhancing thinning efficiency and improving overall fruit quality in apple cultivation.
• The availability of this data promotes further research and development in agricultural automation.