Deep learning based lily bulb classification from multiple camera views

The lily (Lilium spp.) is an important bulbous plant as cut flower species. To produce marketable lily flowers, high-quality lily bulbs are required from growers. Depending on the size, shape, and other characteristics, the bulbs may have varied economic value. The factor that reduces the quality of the bulb and consequently reduces customer satisfaction is the occurrence of “double-nosed” and “defected” bulbs. Recognizing them during the processing is a laborious and time-consuming task. Therefore, an accurate vision system to classify and analyse the bulbs in real time is critical. In this paper, we describe the results of our study on building a deep learning-based model for lily bulb classification. We use a custom-made system for data collection. It comprises 9 cameras for imaging the bulbs from multiple sides. We utilize the pre-trained EfficientNet deep learning model as the backbone and train multi-view convolutional neural network (CNN) on various combinations of camera views. Then, based on the performance, we propose the best combination of camera views to be used in building the sorting machine. The accuracy of 97% was achieved during the experimental study.