摘要
针对高空绝缘子背景复杂、缺陷目标较小等因素造成小目标检测效果不佳,提出了一种改进的基于残差融合注意力模块的SSD(Residual Fusion Attention Module,RFAM-SSD)小目标检测算法。首先将网络一分为二,主干网络通过残差网络的多重特征循环融合模块(ResNet Multiple Cycle Feature Fusion Module,RES-MFCFM)得到特征层,经过卷积得到六个预测特征层,分支网络通过卷积得到对应的六个特征层,两分支通过RFAM,得到最终的六个特征层来检测目标;设计Focal-IOU Loss来代替原损失,提高检测效果。实验表明改进后的算法mAP为92.4%,比原始SSD算法提高了7.2%,且满足实时检测需求,表明该检测算法对于绝缘子缺陷小目标有较好的检测效果。
According to the poor detection effect caused by the complex background of high-altitude insulators and the small defect target,an improved small-target detection algorithm based on Residual Fusion Attention Module(RFAM-SSD)is proposed.Firstly,the network is divided into two parts.The backbone network obtains the feature layer through the Multiple Cycle Feature Fusion Module(ResNet Multiple Cycle Feature Fusion Module,RES-MFCFM)of the residual network,and six predictive feature layers are obtained through convolution.The branch network obtains the corresponding six feature layers through convolution,while the two branches obtain the final six feature layers through RFAM to detect the target.Focal-IOU Loss is designed to replace the original loss and improve the detection effect.Experiments show that the mAP of the improved algorithm is 92.4%,which is 7.2%higher than that of the original SSD algorithm,and meets the real-time detection requirements,indicating that the detection algorithm has a good detection effect on small insulator defect targets.
作者
俞俊
武丽
付相为
张征浩
葛彩成
YU Jun;WU Li;FU Xiang-wei;ZHANG Zheng-hao;GE Cai-cheng(School of Electronics and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210000,China;School of Electronic Information Engineering,Wuxi University,Wuxi 214000,Jiangsu Province,China)
出处
《信息技术》
2024年第12期72-79,共8页
Information Technology
基金
国家青年自然基金项目(62106111)
2021年第二批产学合作协同育人项目(202102563020)。
关键词
小目标检测
SSD算法
残差网络
注意力机制
特征融合
small target detection
SSD algorithm
residual network
attention mechanism
feature fusion
作者简介
俞俊(1995-),男,硕士研究生,研究方向为目标检测、图像处理;通讯作者:武丽(1983-),女,硕士,副教授,研究方向为人工智能、模式识别。
