In a study published by Scientific Reports, a new approach for automating pear-picking, termed the high-level deformation-perception network with multi-object search non-maximum suppression (HDMNet), has been introduced. Utilizing deep learning, this method aims to enhance the visual system for robotic pear harvesting, addressing the challenges faced by current object detection models in complex agricultural environments.
Pears, being among the top five fruits globally in terms of economic and nutritional value, present a considerable challenge in harvesting due to the labor-intensive and time-consuming nature of the task. HDMNet seeks to revolutionize this process by providing a high-precision object detection network capable of overcoming the limitations faced by traditional methods. It is built on the foundation of you only look once version 8 (YOLOv8), but introduces three significant improvements: a high-level semantic focused attention mechanism (HSA), a deformation-perception feature pyramid network (DP-FPN), and a multi-object search non-maximum suppression (MO-NMS) to effectively handle the complexities of pear orchard environments.
The research involved the collection and labeling of a dataset containing 8363 images of pear trees, which was used to train and benchmark the HDMNet model. Results demonstrated that HDMNet outperforms existing models with a mean average precision (mAP) of 75.7%, achieving high efficiency and accuracy in real-time detection tasks. Furthermore, the integration of HDMNet into an Internet of Things (IoT) system for real-world applications was explored, showcasing the model's potential to significantly improve the efficiency and quality of pear harvesting operations.
While the study focuses on pear-picking, the implications of HDMNet extend beyond, offering potential applications in various agricultural tasks requiring high-precision and real-time object detection. This includes the detection of other fruits, crop disease identification, weed identification, and animal tracking, highlighting the model's versatility and potential to contribute to the advancement of agricultural automation and intelligence.
Source: azorobotics.com