YOLO-AGU: A Lightweight Target Detector for Critical Components in Artillery Autoloading Systems

The automatic loading system is an important part of the artillery and an indispensable part of the artillery intelligence process. The complex combat environment of the artillery autoloading chamber environment and the high impact of the combat mission of each component is prone to various failures, so it is necessary to use the visual system to assist in fault diagnosis. Problems such as occlusion and illumination changes make it challenging to monitor the health status of each component of automatic loading in a complex environment. To address these issues, this paper proposes a vision inspection method suitable for harsh imaging environments that can accurately identify critical components within an autoloading chamber in real-time. Based on YOLOv8, this paper introduces a lightweight image preprocessing module, AOD-Net, to perform preliminary enhancement processing on the input image before feature extraction to improve the target detection performance of YOLOv8 under a harsh imaging environment. Meanwhile, GSConv and UIB are incorporated to combine to build an efficient neck, called GU-Neck. In this architecture, GSConv combines the advantages of SC and DSC, and introduces the shuffle operation to uniformly mix the features, which enhances the feature expression ability of the model while maintaining efficient computation; the UIB module can further reduce the computational cost, and the redundant of gradient information transfer and improve the computational efficiency and feature reuse capability. The proposed method is evaluated using a self-constructed data set of magazine structural components. The experimental results show that the YOLO-AGU model can meet the requirements for accurate real-time monitoring of structural components in harsh environments, with a 4% improvement in mAP@50-95 compared with baseline, and the model is lightweight enough to be deployed on an on-board embedded processor.