基于改进SSD的缺陷目标红外检测算法  

An infrared defect target detection algorithm based on improved SSD

作　　者：张华忠[1] 杨荣 邓旭 李飞 钟勉 ZHANG Hua-zhong;YANG Rong;DENG Xu;LI Fei;ZHONG Mian(Institute of Electronic and Electrical Engineering,Civil Aviation Flight University of China,Guanghan 618307,China;Sichuan General Aircraft Maintenance Technology Engineering Research Center,Guanghan 618307,China)

机构地区：[1]中国民用航空飞行学院,航空电子电气学院,四川广汉618307 [2]四川省通用航空器维修工程技术研究中心,四川广汉618307

出　　处：《激光与红外》2024年第12期1885-1893,共9页Laser & Infrared

基　　金：四川省通用航空器维修工程技术研究中心项目(No.GAMRC2021YB12)资助。

摘　　要：在外场实验时,由于民机复合材料蒙皮缺陷红外检测缺陷特征不明显,导致检测精度低和复杂模型导致检测速度慢,针对该问题,提出一种改进的SSD算法提高检测精度和实现模型轻量化。该算法首先采用U-Net网络对图像预处理,降低无关特征信息的干扰,增强缺陷的可检测性。其次,使用Mobilenetv2作为骨干网络,减少模型所占内存大小,提高缺陷检测效率。然后,引入融合改进注意力机制(CBAM)的倒残差模块作为辅助卷积层,进一步轻量化模型并解决精度低的问题。消融实验和对比实验表明,该算法在民机复合材料缺陷数据集上,mAP精度高达96.8%,检测速度(FPS)为72.74 f/s。相比传统SSD算法,AP0.5提升了8.3%,参数量(Params)减少至3.966 M,浮点量(GFLOPS)降低了42倍。该算法在飞机复合材料红外检测领域具有良好的应用前景。In the field experiment,due to the lack of obvious defect characteristics in infrared detection of civil aircraft composite skin defects,resulting in low detection accuracy and complex model leads to the slow detection speed.To solve these problems,an improved SSD algorithm is proposed to enhance the detection precision and realize the model lightweight.Firstly,U-Net network is used for image preprocessing to reduce the interference of irrelevant feature information and enhance the detectability of defects.Secondly,Mobilenetv2 is used as the backbone network to reduce the memory size of the model and improve the efficiency of defect detection.Then,the inverse residual module of Convolutional Block Attention Module(CBAM)serves as an auxiliary convolution layer to further lightweight the model and address precision reduction.The ablation experiments and comparison experiments show that the mAP accuracy of the proposed algorithm is as high as 96.8%on the defect data set of civil aircraft composites,with a detection speed of 72.74 f/s(FPS).Compared with the traditional SSD algorithm,AP0.5 improves by 8.3%,the number of parameters counts(Params)is reduced to 3.966 M,and the number of floating points(GFLOPS)is reduced by 42 times.This algorithm has a good application prospect in the field of infrared detection of aircraft composites.

关 键 词：红外检测 U-Net网络 SSD算法 mAP精度 轻量化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN219[自动化与计算机技术—控制科学与工程]