摘要
深度学习技术可显著提高军事作战中坦克目标识别的准确率和速度,减少误判和漏检,从而降低人员伤亡和战争损失.为了解决大型、复杂、耗时的模型可能会受限于计算资源、存储和能耗等方面的问题,提出了 一种基于YOLOv5轻量化目标检测器.通过C2f based on attention mechanism模块丰富梯度流信息并进一步加速运算;Lead Head与Aux Head的有机结合平衡正负样本提高模型对遮挡小坦克目标识别漏检的能力;利用FasterNet作为特征提取网络,解决了参数量大、算力要求高的问题.实验结果表明:相较于原始YOLOv5,改进后的模型Map0.5、mAP0.5∶0.95分别提高了 1.2％和4.2％,参数量、GFLOPs以及Best.pt分别降低了32.3％、27.59％和26.01％.改进后的YOLOv5模型可以非常快速精准识别坦克目标,通过模型的轻量化使其更容易在移动端和嵌入式设备上部署.
Abstract
The application of deep learning techniques significantly enhances the accuracy and speed of tank target recognition in military operations,reducing misjudgments and omissions,thereby minimizing casualties and war losses.To address the limitations of large,complex,and computationally intensive models in terms of computing resources,storage,and energy consumption,a YOLOv5-based lightweight object detection system was proposed.This approach enriches gradient flow information and further accelerates computations through the C2f module based on attention mechanism.The combination of Lead Head and Aux Head balances positive and negative samples,improving the mod-el's ability to detect obscured small tank targets.Additionally,the utilization of FasterNet as the feature extraction net-work resolves issues related to high parameter quantity and computational demands.Experimental results demonstrate that compared to the original YOLOv5,the improved model achieves a 1.2％and 4.2％increase in MapO.5 and mAP0.5:0.95,respectively,while reducing parameters,GFLOPs,andBest.pt by 32.3％,27.59％,and 26.01％.The improved YOLOv5 model enables fast and accurate tank target recognition,making it more accessible for deploy-ment on mobile and embedded devices due to its lightweight nature.
NETL
NSTL
万方数据