To build a Mask R-CNN based detection model, he took hundreds of photos of apple blossom clusters. He then developed a Kingflower Segmentation algorithm to identify and locate kingflowers from a raw dataset of apple blossom images. The study was conducted at Penn State University’s Center for Fruit Research and Dissemination, Biglerville.
Gala and Honeycrisp Apple varieties were selected for testing. Test trees were planted in 2014 at approximately 5 ft (gara) and 6 1/2 ft (honeycrisp) spacing. These trees were trained in a tall spindle canopy architecture with an average height of about 13 feet. An image acquisition system with a camera was mounted on a utility vehicle that moved between rows of trees.
Mu noted that training a machine vision system to find the king flower is difficult. Because the king flower is the same size, color and shape as the side flowers in the cluster, the king flower is usually hidden by the surrounding flowers due to its central location.
To meet the transfer learning requirements for Mask R-CNN model training, the raw images were labeled with two predefined classes: distinct flowers and occluded flowers. To improve accuracy, Mu explains that they used a data augmentation approach to grow the training dataset four times his.