An automated algorithm for fruit image dataset building

This paper introduces a new algorithm that utilises images from the Fruits-360 dataset, superimposes them of various backgrounds and creates associated annotation files with the coordinates of the bounding boxes surrounding the fruits. The main challenge of this task was accounting for the variations in lighting and occlusion associated with outdoor locations. The utility and application of such an algorithm is to reduce the need to collect real world data for training, accelerating the speed at which new models are developed. Using 3000 images generated by this algorithm we train a single shot multibox detector (SSD) to study the feasibility of using generated data during training. We then test the trained model on 70 real world images of apples (65 images of apples on trees and 5 images of apples in bunches) and obtain a mean average precision of 0.750 and we compare our results with those obtained by other state of the art models.

---

1. Short article

---

2. Track 1: 17th International Symposium on Advanced Artificial Intelligence in Applications

---

3. Opracowanie rekordu ze środków MEiN, umowa nr SONP/SP/546092/2022 w ramach programu "Społeczna odpowiedzialność nauki" - moduł: Popularyzacja nauki i promocja sportu (2022-2023).