多尺度特征增强的噪声鲁棒图像边缘检测算法  

Noise-robust image edge detection algorithm based on multi-scale feature enhancement

作　　者：王富平 任晶晶 公衍超 李藕 刘卫华 张泽均 WANG Fuping;REN Jingjing;GONG Yanchao;LI Ou;LIU Weihua;ZHANG Zejun(School of Communications and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;College of Physics and Electronic Information Engineering,Zhejiang Normal University,Jinhua 321004,China)

机构地区：[1]西安邮电大学通信与信息工程学院,陕西西安710121 [2]浙江师范大学物理与电子信息工程学院,浙江金华321004

出　　处：《西安邮电大学学报》2025年第1期27-36,共10页Journal of Xi’an University of Posts and Telecommunications

基　　金：国家自然科学基金项目(61802305);陕西省重点研发计划项目(2024SF-YBXM-663)。

摘　　要：针对图像受噪声干扰导致边缘特征检测精度下降的问题,提出一种多尺度特征增强的噪声鲁棒图像边缘检测算法。该算法将多分支结构和噪声抑制模块嵌入深度边缘检测基线网络中,在浅层利用多层卷积降采样模块生成多尺度特征图,增强多尺度边缘信息。针对不同尺度的边缘检测,设计多尺度噪声抑制模块,利用增强注意力模块和残差连接实现噪声抑制,并联合噪声抑制损失和边缘检测损失监督模型训练,以实现噪声鲁棒的边缘检测。分别在BIPED数据集和BSDS500数据集上进行检测,在噪声标准差为50时,所提算法在BIPED数据集上的最优数据集尺度为0.848、最优图像尺度为0.857及平均准确率为0.908,均优于对比算法。结果表明,在不同等级噪声的干扰下,所提算法的噪声鲁棒性较好,能够有效提升噪声环境下的边缘检测精度。For the problem that the accuracy of edge feature detection decreases when the image is disturbed by noise,a noise robust image edge detection algorithm based on multi-scale feature enhancement is proposed.The multi-branch structure and noise suppression module are embedded into the deep edge detection baseline network,and the multi-scale convolution downsampling module is used to generate multi-scale feature maps in the shallow layer to enhance multi-scale edge information.For edge detection at different scales,a multi-scale noise suppression module is designed,which uses the enhanced attention module and residual connection to achieve noise suppression.The noise suppression loss and edge detection loss are combined to supervise the model training to achieve noise-robust edge detection.The proposed algorithm is tested on the BIPED dataset and the BSDS500 dataset.When the noise standard deviation is 50,the global optimal thresholds of the proposed algorithm on the BIPED dataset are 0.848 for the optimal dataset scale,0.857 for the optimal image scale and 0.908 for the average accuracy,which are better than those of the comparison algorithms.It shows that the proposed algorithm has good noise robustness under different levels of noise interference,and can effectively improve the edge detection accuracy in noisy environments.

关 键 词：边缘检测 无人机 噪声鲁棒 多尺度特征 特征注意力 

分 类 号：TN911.73[电子电信—通信与信息系统]