A fast and efficient green apple object detection model based on Foveabox

Abstract

Fruit object detection is crucial for automatic harvesting systems, serving applications such as orchard yield measurement and fruit harvesting. In order to achieve fast recognition and localization of green apples and meet the real-time working requirements of the vision system of harvesting robots, a fast optimized Foveabox detection model (Fast-FDM) is proposed. Fast-FDM uses an optimized form of anchor-free Foveabox to accurately and efficiently detect green apples in harvesting environments. Specifically, the EfficientNetV2-S with fast training and small size is used as the backbone network, a weighted bi-directional feature pyramid network (BiFPN) is employed as the feature extraction network to fuse multi-scale features easily and fast, and then the fused features are fed to the fovea head prediction network for the classification and bounding box prediction. Furthermore, an adaptive training sample selection (ATSS) method is adopted to directly select positive and negative samples, allowing green fruits of different scales to obtain higher recall and achieve more accurate green apple detection. Experimental results show that the proposed Fast-FDM realizes a mean average precision (mAP) of 62.3% for green apple detection using fewer parameters and floating point of operations (FLOPs), achieving better trade-offs between accuracy and detection efficiency.