融合图像去噪和特征对齐的目标检测算法  

Target detection algorithm with image denoising and feature alignment

作　　者：许莉 张玉雪 范纯龙 张加金 XU Li;ZHANG Yu-xue;FAN Chun-long;ZHANG Jia-jin(School of Computer Science,Shenyang Aerospace University,Shenyang 110136,China)

机构地区：[1]沈阳航空航天大学计算机学院,辽宁沈阳110136

出　　处：《计算机工程与设计》2025年第2期392-398,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62171295)。

摘　　要：为提高目标检测模型的对抗鲁棒性,提出一种融合图像去噪和特征对齐的目标检测算法(target detection algorithm with image denoising and feature alignment,TDA-IDFA)。采用坐标注意力机制对图像去噪结构进行改进,在去除图像噪声的同时强化图像细节和边缘信息;结合特征对齐对抗训练方法,训练不仅鲁棒而且准确的目标检测模型,在提升目标检测模型对抗鲁棒性的同时,平衡模型对干净样本和对抗样本的识别精度;使用不同的对抗攻击方法,针对多种算法展开对比实验。实验结果表明,所提算法在PASCAL VOC和MS-COCO数据集的对抗鲁棒性比现有算法分别平均提高了9.75%和8.72%。To improve the adversarial robustness of the target detection model,a target detection algorithm combining image denoising and feature alignment(TDA-IDFA)was proposed.The coordinate attention mechanism was used to improve the image denoising structure,which enhanced image details and edge information while removing image noise.Combined with feature alignment adversarial training methods,a robust and accurate target detection model was trained,which improved the robustness of the target detection model against adversarial attacks while balancing the recognition accuracy of the model against clean and adversarial example.Comparative experiments were conducted using various adversarial attack methods against multiple algorithms.Experimental results show that the adversarial robustness of the proposed algorithm on PASCAL VOC and MS-COCO datasets is improved by an average of 9.75%and 8.72%compared to that of existing algorithms,respectively.

关 键 词：目标检测 图像去噪 注意力机制 特征对齐 对抗攻击 鲁棒性 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]