Real-time object detection for screening cannabis seed gender

Abstract

Perception and understanding of cannabis are more expansive than they formerly were. Almost all growers are primarily interested in getting harvests of big flower buds from cannabis female plants since THC, CBD and other cannabinoids are found in female flowers and valuable for medical and industrial market segments. Selecting only female seeds to cultivate is thus an important step to produce THC, CBD profitably. Unfortunately, outdoor cultivation in Thailand traditionally grows regular cannabis seeds that grow up of mixed male and female plants. The male plants will be later spot and eliminated during the pre-flowering stage. This incurs the higher cost of investment and the economic loss consequence. A smart farming approach using AI technology is thus introduced for screening cannabis seed genders before cultivation. A dataset of cannabis seed images of Hang Kra Rog, a well-known Thai cannabis cultivar, was collected from several regions. Data augmentation techniques were carried out to increase the sample size and improve the quality of images. The two object detection models, YOLOv5, were constructed using the initial and augmented datasets. The model trained on the augmented image dataset outperformed the other and achieved the higher precision of 96.4 %, recall of 97.4 %, and mAP_0.5 of 98.7 % with detection speed at 7.2 ms. Moreover, an approach of semi-automated image annotation is also presented in this work. The library of OpenCV is mainly used to facilitate operating various image processing technique to generate the initial image annotation. The preliminary result is promising. The performance of the model using the proposed semi-automated annotation achieved comparable performance to the manual annotation model, and reduced half of the time spent on the annotation process.