面向隐私保护的无镜头成像坐姿识别技术  

Privacy protection oriented lensless imaging sitting posture recognition technology

作　　者：朱斌杰 李裕麒 Zhu Binjie;Li Yuqi(School of Information Science&Engineering,Ningbo University,Ningbo Zhejiang 315000,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315000

出　　处：《计算机应用研究》2025年第4期1262-1267,共6页Application Research of Computers

基　　金：宁波市公益性研究计划资助项目。

摘　　要：现有基于视觉的坐姿识别方法普遍存在过度采集个体生物信息的问题,在追求高识别精度的同时,未充分考虑个人隐私的保护,从而增加了个人信息泄露的风险。针对上述问题,提出了一种在无镜头成像环境下基于层叠特征融合区域注意力增强的坐姿识别方法。该方法旨在利用无镜头成像技术下的模糊图像,通过设计特征融合与锐化模块,结合改进的级联分组注意力机制,增强了模型对关键特征和细节的捕捉能力。同时,采用组合损失函数优化了模型性能。实验结果表明,所提方法在自建无镜头坐姿数据集上,在准确率、精确度、召回率和F 1-score上分别达到了0.96477、0.93196、0.93527和0.93246,均高于其他对比方法,有效提升了坐姿识别的隐私保护性和识别精度。Existing vision-based sitting posture recognition methods generally suffer from excessive collection of individual biometric information.While pursuing high recognition accuracy,they fail to fully consider the protection of personal privacy,thereby increasing the risk of personal information leakage.In response to these issues,this paper proposed a sitting posture recognition method based on cascaded feature fusion and regional attention enhancement in a lensless imaging environment.This method aimed to utilize blurred images obtained through lensless imaging technology.By designing a feature fusion and sharpening module combined with an improved cascaded grouped attention mechanism,it enhanced the model’s ability to capture key features and details.Additionally,it employed a composite loss function to optimize model performance.Experimental results demonstrate that the proposed method,when evaluated on a self-constructed lensless sitting posture dataset,achieves an accuracy,precision,recall,and F 1-score of 0.96477,0.93196,0.93527,and 0.93246,respectively,surpassing other comparative methods.This effectively enhances both privacy protection and recognition accuracy in sitting posture recognition.

关 键 词：坐姿识别 隐私保护 ResNet 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]