An Annotated Image Dataset for Small Apple Fruitlet Detection in Complex Orchard Environments

Fruit thinning is an essential agricultural practice in apple production systems, critically influencing fruit quality optimization, tree load management, and overall orchard profitability. Traditional manual thinning operations face challenges of labor inefficiency and high costs, while computer vision-based automated solutions present a viable technological alternative. However, the advancement of such intelligent systems has been constrained by the lack of specialized, high-quality datasets for detecting small pre-thinning apples with low chromatic contrast against complex canopy backgrounds. To address this gap, this study presents the small apple pre-thinning dataset designed to provide reliable data support for small apple detection and the development of intelligent thinning systems. The dataset comprises 2,519 high-resolution RGB images (original size 3024\u00d73024 pixels, uniformly resized to 500\u00d7500 pixels for standardization) systematically captured under real-world orchard conditions. The dataset encompasses natural variations in weather conditions (sunny\/cloudy), lighting scenarios (direct sunlight\/backlight), and fruit sizes (3-25mm diameter range) to ensure broad applicability. Each image was meticulously annotated using LabelImg software, with all small apple targets precisely labeled using both PASCAL VOC (XML) and YOLO (TXT) format bounding boxes, facilitating compatibility with various detection frameworks. Validation experiments demonstrate the dataset\u2019s effectiveness across multiple state-of-the-art detection architectures (Faster R-CNN, Cascade R-CNN, Grid R-CNN, RetinaNet, and YOLO networks), with all models achieving consistently high detection accuracy under various challenging conditions. The main contributions of this study include: (1) establishing a dedicated dataset for pre-thinning small apple detection; (2) providing high-quality, multi-scene images with fine annotations; and (3) validating the dataset\u2019s performance across different detection models through comparative experiments. This dataset serves as a valuable resource for developing intelligent thinning systems, with potential applications in promoting automation in the apple industry, enhancing thinning efficiency, and improving fruit quality. Future work will focus on expanding the dataset\u2019s scale and coverage of scenarios, as well as exploring its applications in fruit growth prediction and early yield estimation.