【作者】 肖峰；

【导师】 胡辑伟； 李玮；

【作者基本信息】 武汉理工大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 近年来,随着工业4.0时代的到来,工业自动化和智能化水平的不断提高,工业紧固件目标检测成为了提升生产效率、降低人工成本的关键技术。工业紧固件目标检测在自动化装配线、质量检测、零件分类等多个领域中发挥着日益重要的作用。尽管如此,工业紧固件目标检测面临着一个主要挑战:缺乏大规模、高质量的工业紧固件图像数据集。另外由于工业紧固件的类型繁多、形状复杂且在不同生产环境中存在着显著的视觉差异,现有的目标检测算法,无法在满足实时性的同时,较好的对工业紧固件数据集进行目标检测。针对上述问题,本文研究了工业紧固件图像增广及目标检测方法,主要研究工作如下:(1)工业紧固件数据增广方法的研究。鉴于目前工业紧固件数据集在规模和多样性方面的显著不足,本文采纳了一种创新的数据增广策略,通过改进的RDB-CycleGAN网络,能够在无需成对数据的情况下,学习源数据集和目标数据集之间的映射关系,由绘制渲染的图像生成与真实紧固件高度相似的图像,实现数据增广的目的。增加密集残差块模块并且引入结构相似性损失函数,更加注重工业紧固件的特征,提高生成图像的质量,有效的增加数据集的规模,提高了样本的多样性。此外,在数据增广过程中还特别注意了图像质量的控制和样本的实用性评估。(2)基于轻量级网络GhostNet及多尺度特征融合的目标检测算法研究。针对工业紧固件检测中实时性和检测精度较差的问题,采用了轻量化网络GhostNet有效地减少了网络的参数量,降低了计算复杂度,通过多尺度特征融合策略,模型能够更有效地利用各个特征的信息,提高准确性。设计小目标检测层,增强模型对小目标的检测能力。针对工业场景的复杂性,进一步对数据进行增广,提高模型的鲁棒性。(3)工业紧固件小目标检测系统的设计与实现。基于上述研究,设计并实现了一个工业紧固件目标检测系统。该系统包括人员管理、数据增广、图片检测、记录查询等模块。系统能够自动处理工业紧固件图像,快速准确地检测出目标紧固件,并输出检测结果,较好地完成工业紧固件的检测任务。更多还原

【Abstract】 In recent years,with the advent of the Industry 4.0 era,the continuous improvement in the level of industrial automation and intelligence has made industrial fastener object detection a key technology for enhancing production efficiency and reducing labor costs.Industrial fastener object detection plays an increasingly important role in various fields such as automated assembly lines,quality inspection,and part classification.However,industrial fastener object detection faces a major challenge:the lack of large-scale,high-quality industrial fastener image datasets.Additionally,due to the wide variety of industrial fasteners,their complex shapes,and significant visual differences in different production environments,existing object detection algorithms cannot perform well on industrial fastener datasets while meeting real-time requirements.In response to these problems,this thesis investigates industrial fastener image augmentation and object detection methods,with the main research work as follows:(1)Research of industrial fastener data augmentation methods.Given the significant lack of scale and diversity in current industrial fastener datasets,this thesis adopts an innovative data augmentation strategy.Through an improved RDB-CycleGAN network,it learns the mapping relationship between source and target datasets without the need for paired data,generating images highly similar to real fasteners to achieve the purpose of data augmentation.By adding dense residual block modules and introducing a structural similarity loss function,more focus is placed on the features of industrial fasteners,improving the quality of generated images,effectively increasing the dataset’s size,and enhancing sample diversity.Moreover,special attention is paid to the control of image quality and the practicality evaluation of samples during the data augmentation process.(2)Research of an object detection algorithm based on the lightweight network GhostNet and multi-scale feature fusion.To address the issues of real-time performance and poor detection accuracy in industrial fastener detection,the lightweight network GhostNet is used to effectively reduce the number of network parameters and computational complexity.Through a multi-scale feature fusion strategy,the model can more effectively utilize the information of various features,improving accuracy.A small target detection layer is designed to enhance the model’s ability to detect small targets.To cope with the complexity of industrial scenes,data is further augmented to improve the model’s robustness.(3)Design and implementation of an industrial fastener object detection system.Based on the aforementioned research,an industrial fastener object detection system is designed and implemented.This system includes modules for personnel management,data augmentation,image detection,and record inquiry.It can automatically process industrial fastener images,quickly and accurately detect target fasteners,and output detection results,effectively completing the task of industrial fastener detection.更多还原