Abstract
Accurate detection of citrus in the natural orchard is crucial for citrus-picking robots. However, it has become a challenging task due to the influence of illumination, severe shading of branches and leaves, as well as overlapping of citrus. To this end, a Dense-TRU-YOLO model was proposed, which integrated the Denseblock with the Transformer and used UNet++network as the neck structure. First of all, the Denseblock structure was incorporated into YOLOv5, which added shallow semantic information to the deep part of the network and improved the flow of information and gradients. Secondly, the deepest Cross Stage Partial Connections (CSP) bottleneck with the 3 convolutions module of the backbone was replaced by the CSP Transformer with 3 convolutions module, which increased the semantic resolution and improved the detection accuracy of occlusion. Finally, the neck of the original network was replaced by the combined structure of UNet++ feature pyramid networks (UNet++-FPN), which not only added cross-weighted links between nodes with the same size but also enhanced the feature fusion ability between nodes with different sizes, making the regression of the network to the target boundary more accurate. Ablation experiments and comparison experiments showed that the Dense-TRU-YOLO can effectively improve the detection accuracy of citrus under severe occlusion and overlap. The overall accuracy, recall, mAP@0.5, and F1 were 90.8%, 87.6%, 90.5%, and 87.9%, respectively. The precision of Dense-TRU-YOLO was the highest, which was 3.9%, 6.45%, 1.9%, 7.4%, 3.3%, 4.9%, and 9.9% higher than that of the YOLOv5-s, YOLOv3, YOLOv5-n, YOLOv4-tiny, YOLOv4, YOLOX, and YOLOF, respectively. In addition, the reasoning speed was 9.2 ms, 1.7 ms, 10.5 ms, and 2.3 ms faster than that of YOLOv3, YOLOv5-n, YOLOv4, and YOLOX. Dense TRU-YOLO is designed to enhance the accuracy of fruit recognition in natural settings and boost the detection capabilities for small targets at extended ranges.
DownLoad: