枝との位置関係を考慮した洋ナシ花の深層学習検出.

Artificial pollination is a heavy-duty task in pear production. Our research aims at developing a method to detect flowers to be pollinated from camera images in an application of robot vision. Images are taken of flower clusters from a certain distance for determining their position within the tree and of flowers taken from close up for artificial pollination. All images are analyzed by the Faster R-CNN (Regions with Convolutional Neural Networks) model. Branch area information is extracted by a multi-scale filter fom the distant view images. The Faster R-CNN model trained on the pear flower image data with the branch area information outperformed the model without that information. When the Intersection over Union is 0.5, the average precision was 0.747 for the distant view images and 0.939 for the close-up images. [ABSTRACT FROM AUTHOR]