Detection of Small Oranges Using YOLO v3 Feature Pyramid Mechanism

Existing approaches to fruit detection experience difficulty in detecting small fruits with low overall detection accuracy. The reasons why many detectors are unable to handle small fruits better are that fruit data sets are small, and they are not enough to train previous models of YOLO. Further, these models used in fruit detection are initialized by a pre-trained model and then fine-tuned on fruit data sets. The pre-trained model was trained on the ImageNet data set whose objects have a bigger scale than that of the fruits in the fruit pictures. Fruit detection being a fundamental task for automatic yield estimation, the goal is to detect all the fruits in images. YOLO-V3 uses multi-scale prediction to detect the final target, and its network structure is more complex. Thus, in this work, YOLO-V3 is used to predict bounding boxes on different scales and to make multi-scale prediction, thereby making YOLO-V3 more effective for detecting small targets. The feature pyramid mechanism integrates multi-scale feature information to improve the detection accuracy.