VISUAL ONION GROWTH STAGEDETERMINATION USING CNNS

Abstract: With the growing agricultural sector, the demand for more harvest is increasing. Thus next stepin development is automating the agriculture sector. Onions are widely used in various dishes andhold a significant position as a crop of focus for Ekobot. Ekobot uses Red, Green, Blue (RGB)images upon a trained Convolutional Neural Network (CNN) model to distinguish the onions fromweeds and remove them by mechanical arms. This project is a collaboration between students atMälardalen University and Ekobot. The project intends to assess if it is feasible utilising CNNmodels with Ekobots camera systems to identify the onions growing stage according to height andamount of leaves. Collecting datasets containing real onions and plastic onions made from cableties will be used to train the different CNN models. The plastic onions were easier to preprocessand annotate automatically using the Color Index of Vegetation (CIVE) function to segment theonions but were not as good on real onions with overlaying unpredictable leaves. The depth channelon the cameras was thought to improve segmentation as well but was found insufficient, due tothe small size of onions (under 3.5 mm) in width and a camera specification of +/- 5 mm. Thetraining of the CNNs is on the plant leaf amount and height, which presents promising results onthe plastic onions, with 96.31% on individual images. While a mean average length annotated thereal onions, they performed better looking at them in batches in a heatmap rather than individualclassification that demanded an improved annotated dataset.