Jujube-YOLO: a precise jujube fruit recognition model in unstructured environments

Accurate and efficient detection of jujube fruits in unstructured environments is considered a key challenge. The distinction between cracked and crack-free jujube is considered crucial for enabling selective harvesting by robots. The issues of low detection performance and difficulty distinguishing cracked from crack-free jujube fruits are addressed by the proposed Jujube-YOLO model, based on an improved YOLOv11. A double convolution squeeze-and-excitation (DCSE) module is integrated into the backbone network, and a rectangular self-calibration module (RCM) with the C3k2 module is introduced to enhance the expression of initial features and the extraction of multi-scale contextual information from fresh jujube fruits. A multi-branch channel attention (MBCA) module is designed to replace the standard convolution in the neck network, enabling effective fusion of shallow detail, deep semantic, and multi-scale information. The experimental results show that Jujube-YOLO achieves precision, recall, [email protected], and F1 of 98.37 %, 94.96 %, 97.65 %, and 96.63 %, respectively, with performance shown to be superior to that of Faster R-CNN, YOLOv3, YOLOv3-tiny, YOLOv5n, YOLOv6n, YOLOv8n, and YOLOv11n. At the same time, a practical analysis of lighting conditions, occlusions, speed, sample sources, and model size is performed, and it is concluded that Jujube-YOLO is capable of completing the recognition task in unstructured environment The Jujube-YOLO model is designed for recognizing fresh jujube fruits in orchards, offering theoretical insights for quality assessment, growth monitoring, and selective harvesting robots. The code will be released on GitHub. (https://github.com/wangshuheng000210/Jujube-YOLO.git).