Intelligent detection of Multi-Class pitaya fruits in target picking row based on WGB-YOLO network

In a densely planted orchard, factors such as light variation, branch occlusion, and fruit in non-picking rows had a great impact on the pitaya detection accuracy. In this study, a new WGB-YOLO network was developed and tested for multi-class pitaya fruits detection in target picking rows. The proposed WFE-C4 module was obtained by adding two wings feature enhancement structure based on Bottleneck and cascading MetaAconC functions, which independently enhanced feature extraction from the channel and spatial dimensions. A backbone network with WFE-C4 to replace YOLOv3′s Darknet53 was constructed. The proposed GF-SPP used average pooling and global average pooling instead of 2 maximum pooling in SPP, and the global average pooling features were used as independent channels to strengthen the average and maximum pooling features respectively, which simultaneously achieved multi-scale fusion of features and feature enhancement. The new WGB-YOLO network used a Bi-FPN structured head network to achieve a balanced fusion of multi-scale features. The tests showed that the mAP of multi-lass pitaya in the target picking rows was 86.0% using WGB-YOLO detection, while the AP of NO, FCC, and OB fruit were 96.0%, 84.4%, and 77.6%, respectively. WGB-YOLO improved the AP of the original model for detecting OB fruits by 10.5%, which indicated a significant improvement in model detection performance. Compared with 8 other deep networks such as YOLOv7, WGB-YOLO obtained the highest mAP for detecting multi-class pitaya while maintaining a better detection speed. WGB-YOLO showed good performance in detecting pitaya in densely pitaya planted orchards, which provided a technical foundation for fruit detection in robotic picking of the target rows.