特征差异增强与残差蒸馏网络结合的医药可见光图像异物检测  

Foreign object detection in pharmaceutical visible-light images using feature difference enhancement and residual distillation network

作　　者：刘优武 张辉 孔森林 陶岩 李冲 LIU Youwu;ZHANG Hui;KONG Senlin;TAO Yan;LI Chong(School of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha 410114,China;School of Robotics,Hunan University,Changsha 410012,China;Truking Technology Limited,Changsha 410600,China)

机构地区：[1]长沙理工大学电气与信息工程学院,湖南长沙410114 [2]湖南大学机器人学院,湖南长沙410012 [3]楚天科技股份有限公司,湖南长沙410600

出　　处：《智能系统学报》2025年第1期118-127,共10页CAAI Transactions on Intelligent Systems

基　　金：科技创新2030-“新一代人工智能”重大项目(2021ZD0114503);国家自然科学基金重大研究计划项目(92148204);国家自然科学基金项目(62027810);湖南省科技创新领军人才项目(2022RC3063);湖南省十大技术攻关项目(2024GK1010);湖南省重点研发计划项目(2023GK2068,2022GK2011).

摘　　要：医药中的异物通常形态微弱,导致轻量化算法无法准确检测,而高精度算法通常实时性差。为兼顾医药异物检测的实时性与准确性,提出了一种深度学习蒸馏算法,能够快速、准确地检测药液图像中的异物。首先,在教师网络中引入基于语义特征的上采样方法,增强了教师网络与学生网络之间的特征差异。同时,在学生网络的训练图像中加入随机噪声,提高了在高干扰场景下的鲁棒性。为验证算法的有效性,在灯检设备采集了药液异物数据集并进行了对比实验,蒸馏后平均精度提升了4.1百分点,每秒帧数达到了65,优于目前已有的先进方法。最后,在天池酒液数据集进行拓展实验,检测的平均精度提升了3.9百分点,验证了模型在类似场景中的适用性。Foreign objects in pharmaceuticals are typically small,which causes difficulty for lightweight algorithms to detect them accurately,while high-performance algorithms often struggle with real-time capability.To balance real-time performance and accuracy,a deep learning distillation algorithm is proposed for the precise and rapid detection of foreign objects in pharmaceutical liquid images.The teacher network incorporates a semantic feature-based upsampling method to enhance the feature disparity between teacher and student networks.In addition,random noise is added to the training images of the student network to improve robustness in high-noise detection scenarios.To validate the effectiveness of the algorithm,a pharmaceutical liquid foreign-object dataset is collected using lamp inspection equipment,and comparative experiments are conducted.After distillation,the average precision improves by 4.1%,and the model achieves 65 frames per second,which surpasses current state-of-the-art methods.Extended experiments on the Tianchi liquor dataset show a 3.9%improvement in detection accuracy,which demonstrates the applicability of the model in similar scenarios.

关 键 词：医药 异物 轻量化 深度学习 蒸馏 特征差异 上采样 灯检 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]