计算机与现代化 ›› 2023, Vol. 0 ›› Issue (11): 113-119.doi: 10.3969/j.issn.1006-2475.2023.11.018

• 图像处理 • 上一篇    下一篇

基于CP-YOLOX的冷冻电镜图像蛋白质目标检测算法

  

  1. (1.广东工业大学计算机学院,广东 广州 511400; 2.广州生物岛实验室,广东 广州 510700)
  • 出版日期:2023-11-29 发布日期:2023-11-29
  • 作者简介:欧嘉城(1995—),男,广东佛山人,硕士研究生,研究方向:人工智能,图像处理,E-mail: 756784162@qq.com; 曾安(1978—),女,教授,博士后,研究方向:人工智能,数据挖掘,E-mail: zengan2010@126.com; 通信作者:金亮(1979—),男,研究员,博士,研究方向:冷冻电镜成像,E-mail: jin_liang@grmh-gdl.c。
  • 基金资助:
    国家自然科学基金资助项目(61976058, 61772143); 广东省自然科学基金资助项目(2021A1515012300); 广东省科技计划项目(2019A050510041); 广州市科技计划项目(202103000034, 202002020090)

CP-YOLOX-based Algorithm for Protein Target Detection in Cryo-electron Micrographs

  1. (1. School of Computer Science, Guangdong University of Technology, Guangzhou 511400, China;
    2. Guangzhou Biological Island Laboratory, Guangzhou 510700, China)
  • Online:2023-11-29 Published:2023-11-29

摘要: 摘要:针对现有的冷冻电镜图像蛋白质目标检测算法存在特征融合不充分、网络模型复杂、漏选、错选等问题,提出一种冷冻电镜图像目标检测算法(Cryo-Protein YOLOX, CP-YOLOX)。算法主要包含特征提取模块、特征融合模块、输出端。特征提取模块应用本文提出的B-ResBlockX模块,它使用分组的滤波器产生多条特征通道,提高了特征融合能力,从而捕捉更多细节特征。特征融合模块应用本文提出的FastHead模块,它利用多级的扩张卷积模块并且将输出端简化为单通道,可以在不损失精度的情况下,拥有更加轻量的网络结构。同时为进一步提升准确率与收敛速度,位置损失函数加入目标框与预测框的欧氏距离约束。在公开数据集EMPIAR-10028、EMPIAR-10081、EMPIAR-10089上的实验结果表明,对比YOLOX,所提算法的网络参数量仅为5.19×106,mAP(0.5)分别提升了2.4、3.3和2.5个百分点。

关键词: 关键词:冷冻电镜图像, 目标检测, 蛋白质颗粒检测, 轻量化

Abstract: Abstract: A cryo-electron micrograph target detection algorithm (Cryo-Protein YOLOX, CP-YOLOX) is proposed for the existing cryo-electron micrograph protein target detection algorithm with inadequate feature fusion and complex network model, missed detection and false detection. The algorithm mainly contains feature extraction module, feature fusion module, and output side. The feature extraction module applies the B-ResBlockX module proposed in this paper, which uses grouped filters to generate multiple feature channels to improve the feature fusion capability and capture more detailed features. The feature fusion module applies the FastHead module proposed in this paper, which uses multilevel dilated convolution module for feature fusion and simplifies the output to a single channel, which can have a more lightweight network structure without losing accuracy. In order to further improve the accuracy and convergence speed, the position loss function is added with the Euclidean distance constraint between the target frame and the prediction frame. Experimental results on public datasets EMPIAR-10028, EMPIAR-10081, and EMPIAR-10089 showed that the number of network parameters of the proposed algorithm was only 5.19×106, and the mAP(0.5) was improved by 2.4, 3.3 and 2.5 percentage points, respectively, compared with YOLOX.

Key words: Key words: cryo-electron micrographs, target detection, protein particle detection, lightweight

中图分类号: