基于YOLOX的轻量化目标检测算法及其应用
2025-08-07柴炜朕王朝立孙占全
中图分类号:TP391 文献标志码:A
Lightweight object detection algorithm based on YOLOX and its application
CHAI Weizhen, WANG Chaoli, SUN Zhanquan (School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China)
Abstract: Object detection algorithms are widely used in the field of production safety. To address the problems of slow detection speed and low detection accuracy in complex construction environments, an improved YOLOX detection algorithm was proposed. First, based on the lightweight convolution module Ghost moudle, the backbone network was reconstructed to compress the model parameters and computational complexity, thereby improving detection speed. Second, embedding the coordinate attention mechanism at the output of the backbone network to enhance the model's ability to learn key position information. Finally, recursive gated convolution was introduced into the neck network to enhance the model's spatial position perception ability and capture long-range dependencies in the image. The improved model are experimentally validated on the Pascal VOC and SHWD datasets, comparing with the baseline model, mean average precision increase by 1.69% and 1.1% respectively, model parameter count decrease by 18.8% , computational load decrease by 23.3% , and frame rate increase by 7.6% . Deploying the purposed model on terminal devices can be applied to real-time monitoring and detection in construction environments.
Keywords: object detection; lightweight; coordinate attention; recursive gated convolution
目标检测算法在生产安全领域应用广泛,它可以代替传统的人工监督方法,节约人力物力。在智慧工地建设中,将目标检测算法应用于工人安全帽佩戴检测,可以有效保障复杂施工环境下工人的生命安全。但在实际应用场景中,对算法的实时性和检测精度都有较高的要求。
传统的目标检测方法需要人工设计图像特征提取算子再进行检测,代表算法有HOG[1和DPM[2],这种方法过程繁琐,模型通用性和鲁棒性较差,检测准确率较低。深度学习方法不用人为设置图像特征,可以根据图像本身学习其内在的属性,基于深度学习的目标检测算法已被广泛应用于生产安全领域,例如安全帽检测、人脸识别、火焰检测等。目前,基于深度学习的目标检测算法可大致分为2类:一类是两阶段(two-stage)目标检测算法,其代表有Fast R-CNN[3]、Faster R-CNN[4];另一类是单阶段(one-stage)的目标检测算法,其代表有YOLO[5]、SSD和ATSS[等。两阶段检测算法的思路,是将可能存在物体的区域视为候选区域(region proposal),在此候选区域的基础上完成分类和回归任务。这种方法的网络结构复杂、计算量大,检测速度与单阶段检测算法差距很大,不能满足安全帽检测这一实时应用场景的需求。而在单阶段的检测算法中,省去了候选区域的提取工作,经过单阶段的网络直接完成分类和回归任务,因此其检测速度大大提高,可以满足实时检测应用场景的需求。在单阶段检测算法中,YOLO系列检测方法既可以保持较高的精度,又有很快的检测速度,因此在安全帽检测中应用最为广泛。……
