【作者】 吴杰；

【导师】 王磊；

【作者基本信息】 海南大学 ， 人工智能， 2024， 硕士

【摘要】 目标检测是计算机视觉的一个重要领域,其核心任务是确定目标的精确位置并进行准确分类。人脸口罩检测作为目标检测的一个实际应用场景,可用于身份认证、安全识别、防疫监管、社会公共安全与监控等领域。尽管当前大部分的人脸口罩检测研究都是基于深度学习的检测算法,但其检测精度和效率仍有待提高。因此,本文从实际应用的角度出发,基于深度学习的卷积神经网络算法,选取了YOLO(You Only Look Once)系列算法中的YOLOv3、YOLOv7、YOLOv7-Tiny进行研究与改进,并将其应用于人脸口罩检测的实际问题。主要研究内容如下:(1)针对YOLOv3算法在人脸口罩检测场景中的检测精度较低的问题,研究并改进YOLOv3目标检测算法。本文首先基于YOLOv3算法,分析其主要算法原理与结构;其次,通过设计新的网络结构、加入注意力机制、使用完全交并比(Complete Intersection over Union,CIoU)优化算法损失函数、应用Mosaic数据增强,以此改进YOLOv3算法;最后,进行实验,并分析上述方法对算法性能的影响。结果表明,本文改进的YOLOv3算法可有效检测出人脸是否佩戴了口罩,且平均精度均值(mean Average Precision,m AP)较原始YOLOv3算法提高了23.28%,证明了本文方法的有效性。(2)针对希望实现更高检测精度的需求,研究并改进YOLOv7目标检测算法。本文首先基于YOLOv7算法,分析其主要算法原理与结构;其次,通过设计新的网络结构、加入注意力机制、使用高效交并比(Efficient Intersection over Union,EIoU)优化算法损失函数、应用Mosaic和Mixup数据增强,以此改进YOLOv7算法;最后,进行实验,并分析上述方法对算法性能的影响。结果表明,本文改进的YOLOv7算法可有效检测出人脸是否佩戴了口罩,且m AP较原始YOLOv3算法提高了28%,较原始YOLOv7算法提高了3.05%,证明了本文方法的有效性。(3)针对常规算法参数量大、模型复杂度高等问题,以及顺应当前算法轻量化的趋势,研究并改进YOLOv7-Tiny目标检测算法。本文首先基于YOLOv7-Tiny算法,分析其主要算法原理与结构;其次,通过设计新的网络结构、加入注意力机制、使用智慧交并比(Wise Intersection over Union,WIoU)优化算法损失函数,以此改进YOLOv7-Tiny算法;最后,进行实验,并分析上述方法对算法性能的影响。结果表明,本文改进的YOLOv7-Tiny算法在参数量与模型复杂度得到有效降低的同时,可有效检测出人脸是否佩戴了口罩,且m AP较原始YOLOv7-Tiny算法提高了1.16%,证明了本文方法的有效性。综上,本文结合目标检测与人脸口罩检测任务发展现状,提出基于深度学习的人脸口罩检测改进算法,可在不同环境下进行有效地检测。本研究为人脸口罩检测算法提供了一种新的思路与参考,应用前景广阔。更多还原

【Abstract】 Object detection is an important field of computer vision,and its core task is to determine the precise position of objects and classify them accurately.Face mask detection,as a practical application scenario of object detection,can be used in identity authentication,security recognition,epidemic prevention supervision,social public safety and surveillance.Although most of the current research on face mask detection is based on deep learning detection algorithms,its detection accuracy and efficiency still need to be improved.Therefore,from the practical application point of view,this thesis selects YOLOv3,YOLOv7,YOLOv7-Tiny in YOLO(You Only Look Once)series algorithms to study and improve,and applies them to the practical problem of face mask detection.The main research contents are as follows:(1)In order to solve the problem of low detection accuracy of YOLOv3 algorithm in face mask detection scene,YOLOv3 object detection algorithm is studied and improved.This thesis firstly analyzes the main algorithm principle and structure based on YOLOv3algorithm;secondly,YOLOv3 algorithm is improved by designing new network structure,adding attention mechanism,optimizing algorithm loss function by CIoU(Complete Intersection over Union)and applying Mosaic data enhancement;finally,experiments are carried out to analyze the impact of the above methods on algorithm performance.The results show that the improved YOLOv3 algorithm can effectively detect whether the face mask is worn,and the m AP(mean average precision)is increased by 23.28%compared with the original YOLOv3 algorithm,which proves the effectiveness of the proposed method.(2)In order to achieve higher detection accuracy,YOLOv7 object detection algorithm is studied and improved.This thesis firstly analyzes the principle and structure of YOLOv7algorithm based on YOLOv7 algorithm;secondly,YOLOv7 algorithm is improved by designing new network structure,adding attention mechanism,optimizing algorithm loss function by EIoU(Efficient Intersection over Union),applying Mosaic and Mixup data enhancement;finally,experiments are carried out and the influence of above methods on algorithm performance is analyzed.The results show that the improved YOLOv7 algorithm can effectively detect whether the face mask is worn,and the m AP is improved by 28%compared with the original YOLOv3 algorithm and 3.05%compared with the original YOLOv7 algorithm,which proves the effectiveness of the proposed method.(3)In order to solve the problem of conventional algorithms have large parameters and high model complexity,as well as to comply with the current trend of lightweight algorithms,YOLOv7-Tiny object detection algorithm is studied and improved.This thesis firstly analyzes the main algorithm principle and structure based on YOLOv7-Tiny algorithm;secondly,YOLOv7-Tiny algorithm is improved by designing new network structure,adding attention mechanism,and optimizing algorithm loss function by WIoU(Wise Intersection over Union);finally,experiments are carried out to analyze the impact of the above methods on algorithm performance.The results show that the improved YOLOv7-Tiny algorithm can effectively detect whether the face mask is worn while the parameters and model complexity are effectively reduced,and the m AP is increased by1.16%compared with the original YOLOv7-Tiny algorithm,which proves the effectiveness of the proposed method.In summary,this thesis combines the current development status of object detection and face mask detection tasks and proposes an improved face mask detection algorithm based on deep learning,which can perform effective and real-time detection in different environments.This research provides a new idea and reference for the face mask detection algorithm,and has broad application prospects.更多还原