Detection and classification of opened and closed flowers in grape inflorescences using Mask R-CNN

Accurate measurements of the change in total flower count, and the ratio of opened to closed flowers per inflorescence with time, play an important role in studying phenological changes of inflorescences over time. The duration of flowering has an important role in the resulting fruitset and yield. Automation of the flower counting process with inflorescence images, using image processing and morphological tools, is a challenging problem. This is because it involves the processing of images with varying image qualities, and also because of the close similarity in images between the two classes of interests, opened and closed flowers. Our aim is to build a system with one of the most promising deep learning object detection networks, Mask R-CNN, to detect the individual instances of the above two classes separately using the images with no prior alterations. The system should be tested with the images taken with different illumination levels, different backgrounds, and with different scales. Our system was tested with images taken in three consecutive flowering seasons (2018, 2019 and 2020) and showed promising results. These tests also highlighted areas that can be improved to ensure better accuracy.