A high-precision detection method of hydroponic lettuce seedlings status based on improved Faster RCNN

Abstract

In order to improve the efficiency and reduce high cost for seedlings sorting in the raising process of hydroponic lettuce seedlings, we propose an automatic detection method for hydroponic lettuce seedlings based on improved Faster RCNN framework, taking the dead and double-planting status of seedlings growing in a single hole as our research objects. Since the characteristics of hydroponic lettuce seedlings are dense and small in the images, our model uses High Resolution Network (HRNet) as the backbone network for image feature extraction so as to obtain reliable and high- resolution feature expressions. Besides, we adopt focal loss as the classification loss in the Region Proposal Network (RPN) stage to address the imbalance between difficult and easy samples in seedlings classification. We also employ the Region of Interest (RoI) Align instead of the RoI Pooling layer to improve the detection accuracy of seedlings in the different status. The results show that the mean average precision of our method for the hydroponic lettuce seedlings is 86.2%, which is higher than RetinaNet, SSD, Cascade RCNN, FCOS and other detectors. Compared with different feature extraction networks, the detection accuracy of adopting HRNet performs nicely. Therefore, our method presented for the detection of hydroponic lettuce seedlings status can achieve high accuracy and identify seedlings in a problematic status well, which will provide technical support for automatic seedlings detection of hydroponic lettuce.