基于三目注意力机制的高效轻量化麦穗检测算法  

Efficient and lightweight wheat ear detection algorithm with trinocular attention mechanism

作　　者：董燕[1] 刘运东 李卫杰 刘洲峰[1] 李春雷[1] DONG Yan;LIU Yun-dong;LI Wei-jie;LIU Zhou-feng;LI Chun-lei(School of Electrical and Information,Zhongyuan University of Technology,Zhengzhou 450007,China)

机构地区：[1]中原工学院电子信息学院,河南郑州450007

出　　处：《计算机工程与设计》2023年第7期2062-2069,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(U1804157、62072489、61772576);河南高校科技创新团队基金项目(21IRTSTHN013)。

摘　　要：针对复杂大田环境下基于卷积神经网络的麦穗检测方法实时性差、检测精度低的问题,提出一种基于三目注意力机制的高效轻量化麦穗检测算法。采用MobileNetV3作为主干网络对特征进行高效提取;通过融合通道、空间和位置的特征信息设计三目注意力机制,提升模型对关键特征的敏感度;基于条件卷积设计全局多头自注意力机制,增强全局特征的提取能力;选择CIOU作为边框回归损失函数,增强遮挡及重叠目标的检测效果。实验结果表明,与其它优秀的算法相比,所提算法在减少模型容量的同时,提升了检测精度和速度。To address the problems of poor performances in real-time and detection accuracy of wheat ear detection method based on convolutional neural network in complex field scenes,an efficient and lightweight wheat ear detection algorithm with trinocular attention mechanism was proposed.MobileNetV3 was adopted as the backbone network for efficient feature extraction.The trinocular attention mechanism was proposed by fusing channel,space,and location feature information to improve the sensitivity of key features.The global multi-head self-attention mechanism was proposed based on conditional convolution to enhance the ability of global feature extraction.CIOU was chosen as the border regression loss function to improve the detection performance for the occlusive and overlapping targets.Experimental results show that,compared with other well-known algorithms,the proposed algorithm improves the detection accuracy and speed while reducing the model capacity.

关 键 词：复杂场景 麦穗检测 深度学习 轻量化 三目注意力机制 全局多头自注意力机制 条件卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]