Infield Honeysuckle Detection Based on Improved YOLOv5s Under Natural Lighting

Honeysuckle detection through machine vision is an important task for picking honeysuckle timely and efficiently by the picking robot. In this study, image-based object detection methods, including image processing, machine learning (ML), and deep learning (DL), were used to detect honeysuckles in the field to support the vision system of honeysuckle picking robot. A handheld low-cost RGB camera was used to capture 2242 original images containing honeysuckle flowers in different light, angles, and scales. The dataset was expanded to 20,178 using data augmentation where image datasets of 14,747 and 5,431 were used for training and validation, respectively. Three honeysuckle detection models were constructed based on image processing, ML, and DL. The image processing method segments the honeysuckle from the background by color thresholding, hole filling, and corrosion expansion. ML (support vector machine, SVM) and DL (YOLOv3-tiny, YOLOv5s, YOLOv6s, YOLOv7-tiny, and YOLOv8s) models were trained and validated. To solve the difficulties in detection of small targets of honeysuckle, and the imbalance of positive and negative samples, the original YOLOv5s model was improved by fusing the three modules to get an improved YOLOv5s model. The experimental results showed that the overall performance of the improved YOLOv5s model outperformed that of the other methods, with a smaller model size (18.4 MB) and a higher average precision (86.88%). The detection precision was 88.97%, the F1 value was 84.14%, and the detection speed was more than 30 fps. This study demonstrated that the improved YOLOv5s model can be used as a real-time, reliable, and robust detection method for honeysuckle in the field.