Real-Time Detection of Winter Jujubes Based on Improved YOLOX-Nano Network

Achieving rapid and accurate localization of winter jujubes in trees is an indispensable step for the development of automated harvesting equipment. Unlike larger fruits such as apples, winter jujube is smaller with a higher density and serious occlusion, which obliges higher requirements for the identification and positioning. To address the issues, an accurate winter jujube localization method using improved YOLOX-Nano network was proposed. First, a winter jujube dataset containing a variety of complex scenes, such as backlit, occluded, and different fields of view, was established to train our model. Then, to improve its feature learning ability, an attention feature enhancement module was designed to strengthen useful features and weaken irrelevant features. Moreover, DIoU loss was used to optimize training and obtain a more robust model. A 3D positioning error experiment and a comparative experiment were conducted to validate the effectiveness of our method. The comparative experiment results showed that our method outperforms the state-of-the-art object detection networks and the lightweight networks. Specifically, the precision, recall, and AP of our method reached 93.08%, 87.83%, and 95.56%, respectively. The positioning error experiment results showed that the average positioning errors of the X, Y, Z coordinate axis were 5.8 mm, 5.4 mm, and 3.8 mm, respectively. The model size is only 4.47 MB and can meet the requirements of winter jujube picking for detection accuracy, positioning errors, and the deployment of embedded systems.