Preeti
The content discusses the acquisition of a dataset for automated pear picking, highlighting the challenges faced due to limited existing data. A total of 8363 images and 56,406 labels of fruit and pears were collected using a Hikvision camera, covering various postures, positions, and environments. The dataset was standardized and converted to YOLO format for object detection. The YOLOv8 architecture is then introduced, focusing on its improvements over YOLOv5 and its application in automated pear picking tasks. A new method called HDMNet is proposed based on YOLOv8 advancements to enhance object detection in pear picking scenarios.
The High-level Semantic focused attention mechanism module is detailed, aiming to improve feature extraction by highlighting key pear characteristics. The Deformation-perception feature pyramid network (DP-FPN) is introduced to enhance feature fusion and extraction. The Multi-object search non-maximum suppression (MO-NMS) algorithm is proposed as a post-processing technique to filter redundant detections in complex scenarios like automated pear picking.
The MO-NMS algorithm is compared to Soft-NMS, emphasizing its global perspective and effectiveness in handling small objects and clustered scenes. The content concludes by highlighting the significance of MO-NMS in improving detection accuracy and reducing false positives in pear picking automation. The proposed method aims to enhance the performance of object detection systems in challenging environments like automated pear picking.
Source link
Source link: https://www.nature.com/articles/s41598-024-65750-6
GIPHY App Key not set. Please check settings