Attention-Guided Oriented Bounding Box Real-Time Image Analytics Model for Dense Wheat Imperfect Grains and Impurities

Impurities and imperfect grains in dense wheat are critical factors affecting wheat quality and secondary production. Despite the significance, detecting dense wheat grains still faces challenges, mainly due to poor detection accuracy and severe leakage. These issues hinder the model from meeting the reliability and real-time detection requirements in practical scenarios. To address these challenges, we propose an attention-guided oriented bounding box network (AG-OBBNet) for real-time detection of dense wheat imperfect grains and impurities. Firstly, the oriented bounding box method(OBB) is integrated to introduce the orientation of wheat grains into the YOLOX framework, which dramatically improves the accuracy of image analysis and result information visualization. Next, we construct a novel strong feature extraction network that integrates low-dimensional localization features with semantic features enhanced by DS-CSP and a global attention mechanism before fusion. This network facilitates deep mining of semantic information, improving wheat grain classification, reducing computational overhead, and enhancing the utilization of low-dimensional localization features to improve the leakage detection problem in dense wheat grains. Finally, we introduce the Focal-EIOU loss function for faster convergence speed and better localization effects, replacing the original IOU loss function to further enhance model detection performance. To validate the effectiveness and generalization ability of the model, we construct a dense wheat grain dataset (DWGD) categorized into 12 different classes. Experimental results indicate an average accuracy (mAP) of 96.4% for detecting imperfections and impurities in wheat grains, with the Frame Per Second(FPS) of 148.6. Additionally, the model features fewer parameters and meets real-time images analytics and visualization requirements in practical scenarios.