一种改进RetinaNet深度学习网络的城乡结合部违章建筑检测方法  

A Robust Detection Method for Illegal Buildings in Urban-Rural Fringe Areas Based on Improved RetinaNet Deep Learning Network

作　　者：李作进 曹亚男 贺学乐 李明虹 李东阳 Simon X.Yang LI Zuojin;CAO Yanan;HE Xuele;LI Minghong;LI Dongyang;Simon X.YANG(School of Electronic and Electrical Engineering,Chongqing University of Science and Technology,Chongqing 401331,China;School of Engineering,University of Guelph,Guelph Ontario N1G2W1,Canada)

机构地区：[1]重庆科技大学电子与电气工程学院,重庆401331 [2]圭尔夫大学工程学院,安大略圭尔夫N1G2W1

出　　处：《重庆科技大学学报（自然科学版）》2024年第4期75-82,共8页Journal of Chongqing University of Science and Technology(Natural Sciences Edition)

基　　金：国家自然科学基金项目“模糊循环神经网络和驾驶人疲劳特征空间机理研究”(61873043);重庆市自然科学基金项目“面向多模态异构大数据的特征自主学习方法研究”(CSTC2021YCJH-BGZXM0071);重庆市教委科技重大项目“山地道路疲劳驾驶特征融合与险态行为识别研究”(KJZD-M202301502)。

摘　　要：针对城乡结合部违章建筑类型多样化、小目标待检物较多且检测精度较低等问题,提出了一种改进RetinaNet深度学习网络的城乡结合部违章建筑检测方法。首先,在主干特征提取网络ResNet50中嵌入CA注意力机制,以增强网络对小目标的感知能力;其次,在ResNet50中引入并行空洞卷积(DCB)模块,以实现多尺度特征融合,降低算法的漏检率;最后,将ResNet50网络的ReLU激活函数替换为GeLU激活函数,以加快模型收敛,提升模型稳定性。实验结果表明,改进后模型的平均准确率达到93.28%,参数量为3.920×107,可为城乡结合部违章建筑实时监测及拆除工作提供部分理论依据。Aiming at the problems such as diverse types of illegal buildings,a large number of small objects to be detected and low detection accuracy in urban-rural fringe areas,etc.,a detection method for illegal buildings based on improved RetinaNet deep learning network is proposed.Firstly,the CA attention mechanism is embedded in the backbone feature extraction network ResNet50 to enhance the network′s perception of small objects.Secondly,the parallel dilated convolution(DCB)module is introduced into ResNet50 to achieve multi-scale feature fusion and reduce the algorithm′s missed detection rate.Finally,the activation function is replaced by GeLU to accelerate model convergence and improve model stability.Experimental results show that the average accuracy of the improved model reaches 93.28%,and the number of parameters is 3.920×107,which can provide a theoretical basis for the real-time monitoring and demolition of illegal buildings in urban-rural fringe areas.

关 键 词：深度学习 RetinaNet 注意力机制 违章建筑 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]