Green Sweet Pepper Fruit and Peduncle Detection Using Mask R-CNN in Greenhouses

In this paper, a mask region-based convolutional neural network (Mask R-CNN) is used to improve the performance of machine vision in challenging task detecting peduncles and fruits green sweet peppers (Capsicum annuum L.) greenhouses. One most complicated stages pepper harvesting process achieve precise cut peduncle or stem because type specialty crop cannot be grabbed pulled by fruit since integrity value product are compromised. Therefore, accurate detection becomes vital for autonomous peppers. ResNet-101 combined with feature pyramid (FPN) architecture (ResNet-101 + FPN) adopted as backbone extraction object representation enhancement at multiple scales. Mask images generated, focused on pepper, which complex color variety due its resemblance background. addition bounding boxes, R-CNN provides binary masks result instance segmentation, would help localization 3D space, next phase peppers, it isolates pixels belonging demarcates boundaries. The prediction results 1148 100 test showed precision rate 84.53%. 265 71.78%. mean average an intersection over union 50 percent (mAP@IoU=50) model-wide segmentation was 72.64%. time using high-resolution 1.18 s. experimental show that proposed implementation manages segment real-time unmodified production environment under occlusion, overlap, light variation conditions effectiveness not previously reported simultaneous 2D models pepper.