基于注意力机制的食物识别与定位算法  

Food Recognition and Location Algorithm Based on Attention Mechanism

作　　者：彭耿 刘宁钟[1] PENG Geng;LIU Ning-zhong(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106

出　　处：《计算机技术与发展》2022年第11期121-126,共6页Computer Technology and Development

基　　金：国家自然科学基金(61375021)。

摘　　要：随着食物检索、食物推荐和食物监测等应用需求的增加,食物的自动分析成为了研究的热点。由于食物种类多,存在类间差异小、类内差异大、多尺度等特点,食物识别和定位的准确率一直不高。并且目前很多研究,在食物分析任务中,推理速度慢,性能不佳。针对这些问题,结合注意力机制,提出了一个更优的主干网络,能更好地提取食物细粒度特征。同时对Neck部分进行研究,进行多尺度特征融合,提出了一种轻量级的端到端食物识别和定位框架FFAM(Feature Fusion of Attention Mechanism)。在目前具有挑战性的公开数据集UNIMIB2016上的实验结果表明,该算法比目前的很多方法在精度上更具有优势,最终mAP达到了94.1%。由于得到的模型相比YOLOv4精度高且更小,在应对移动端、嵌入式设备中部署食物分析模型解决实际任务时,能有一个更好的性能表现。With the increasing demand for applications such as food retrieval,food recommendation and food monitoring,automatic analysis of food has become a hot research topic.The accuracy of food identification and localization has been low due to the large number of food types with small inter-class differences,large intra-class differences,and multiple scales.And many current studies,in the food analysis task,have slow inference speed and poor performance.To address these problems,a better backbone network is proposed to extract food fine-grained features better by combining the attention mechanism.The Neck part is also investigated for multi-scale feature fusion,and a lightweight end-to-end food identification and localization framework FFAM is proposed.The experiments on the current challenging public dataset UNIMIB2016 show that the proposed algorithm is more competitive than many current methods in terms of accuracy,with a final mAP of 94.1%.Since the obtained model is more accurate and smaller compared to YOLOv4,it can have a better performance when dealing with the deployment of food analysis models in mobile and embedded devices to solve practical tasks.

关 键 词：食物识别与定位 深度学习 注意力机制 特征融合 YOLO 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]