Fruit Detection from Digital Images Using CenterNet

Zhao, Kun; Yan, Wei Qi

In this paper, CenterNet is chosen as the model to settle fruit detection problem from digital images. Three CenterNet models with various backbones were implemented, namely, ResNet-18, DLA-34, and Hourglass. A fruit dataset with four classes and 1,690 images was established for this research project. By comparing those models, followed the experimental results, the deep learning-based model with DLA-34 was selected as the final model to detect fruits from digital images, the performance is excellent. In this paper, the contribution is that we deploy a model based on CenterNet for visual object detection to resolve the problem of fruit detection. Meanwhile, there are 1,690 images distributed in four classes. Throughout evaluating the performance of the model, we eventually affirm the CenterNet based on DLA-34 to detect multiclass fruits from our images. The performance of this method is better than the existing ones in fruit detection.