IMAGE-BASED CANOLA SEED DETECTION AND COUNTING USING YOLOV8N AND YOLO11N OBJECT DETECTION MODELS

Abstract

Accurate seed counting is a key component in seed quality assessment, breeding programs, germination and crop productivity analysis, and counting canola (Brassica napus L.) seeds is essential. The traditional methods of seed counting are time-consuming, labor-intensive, and subject to human errors, particularly in the processing of large quantities of samples. The development of computer vision and deep learning techniques has been seen as a very promising approach to the automated recognition and counting of objects in agriculture in recent years. This paper proposes an image-based automatic canola seed detection and counting method based on YOLOv8n model and YOLO11n models. A customized dataset of 833 images of canola seeds in natural light was created. The images in the dataset comprised different numbers of seeds (from one to ten). Each image was annotated by hand in the Computer Vision Annotation Tool (CVAT) with only one class, “canola_seed”. The dataset was split into 666 train images and 167 validation images. Both YOLOv8n and YOLO11n models were trained for 50 epochs and evaluated using F1-score, mean average precision (mAP@50), and confusion matrix analysis. The results on the experimental set indicate that the YOLOv8n model had a mAP@50 of 0.965 and an optimal F1 score of 0.95, and the YOLO11n model had a mAP@50 of 0.969 and an optimal F1 score of 0.97. YOLOv8n and YOLO11n had mAP@50-95 of 0.270 and 0.266, respectively. Moreover, YOLO11n had fewer false-positive and false-negative results compared with YOLOv8n, demonstrating the reliability of the detection and counting process. The results show that YOLO11n was superior to automated canola seed detection and counting. The proposed framework is a solution that can be used for seed quality testing, agricultural phenotyping, and future smart agriculture applications, which is low-cost, accurate, and efficient.