Efficient target detection algorithm for real-time industrial surveillance based on higher-order information fusion

Abstract

Industrial surveillance target detection presents a multitude of challenges due to their complex environments, numerous operational equipment, and vast discrepancies in the scale of detection targets. These factors collectively contribute to the elevated difficulty of detection in such settings, potentially leading to increased safety hazards during industrial operations. To solve these problems of detection in an industrial complex scenarios, this study proposes a detection model based on YOLO11 called YOLO-HF. To improve the detection of hidden small targets, the MultiPath Coordinate Attention (MPCA) mechanism and Deformable Convolution (DCN) are used in C3k2 module to enhance the spatial sampling position in the module by learning offsets. Then, for the purpose of solving the problem of difficult detection in the case of a large gap in the target size, the method designs a new neck Higher-order information Fusion Path Aggregation Network (HF-PAN). The Depthwise with CARAFE Up-convolution (DCU) block has been developed to improve the feature fusion and computational performance of the model. In order to solve the problems of insufficient existing industrial scene dataset and a big difference to the real scene, the Rootcloud industrial scene dataset is self-proposed, and experimental comparisons are carried out on this dataset. Experimental data illustrate that, compared with the original algorithm, the improved network enhances mAP@0.50 by 4.2% and mAP@0.50:0.95 by 3.7%, and achieves a high frame rate of up to 151 FPS, which effectively enhances the performance of detecting targets in complex scenes and maintains good detecting efficiency.