Detection Mature Bud for Daylily Based on Faster R-CNN Integrated With CBAM

The daylily (Hemerocallis citrina Baroni) is rich in not only nutrition ingredients but also functional components, and the edible part is the flower, not containing its pedicel. The primary challenge in developing a robotic daylily harvester is recognizing mature bud in the unstructured and uncertain environment. The objective of this study is to propose an accurate detection model. Hemerocallis citrina cv. ‘DatongHuanghua’ variety is used in this study. We initially adopt VGG16, VGG19, ResNet50, ResNet101 and ResNet152 as the backbones of Faster R-CNN respectively to build different detection models. The experimental results show that VGG19 and ResNet50 are two best-performing models in the corresponding VGGNet and ResNet, and the Average Precision (AP) of VGG19 is 90.18%, while ResNet50 is 88.35%. Based on these, we further integrate Convolutional Block Attention Module (CBAM) in Faster R-CNN with three different integration modes: plugging CBAM behind Conv5_x of VGG19 and ResNet50 respectively, as well as between every two “bottleneck” blocks of ResNet50. The comparison demonstrate plugging CBAM between every two blocks of ResNet50 is the best integration mode, and the corresponding detection model has a 2.22% highest increase in AP. Therefore, we empirically validate the performance of detection model for daylily mature bud based on Faster R-CNN integrated with CBAM.