Maturity classification of blueberry fruit from camera image for cultivation support system

In this paper, a method for maturity classification of blueberry fruit from camera images is proposed to realize the cultivation support system. According to the methods commonly used in the agricultural field, the maturity of blueberry fruit is classified into five levels proposed by Shutak. To classify the maturity level of blueberry fruit, we first detect the blueberry fruit region from the whole image, and then classify each region into five levels. Our method achieves high accuracy by extracting the local region of each fruit that includes the background and identifying its maturity level by the Transformer, which can learn the relationships not only between inside the fruit region but also outside the fruit. To train and validate the proposed method, the dataset of blueberry fruit image was created. We took images of blueberry fruits with the color chart corresponding to the maturity level of the fruit and labelled for the object detection labels and five maturity levels compared to the color chart. In experiments, we trained several deep learning models to compare with the proposed method. Three experimental patterns with changing combinations of training and test data in the dataset resulted in an average of 93.2% in the maturity classification task by the proposed method.