作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (7): 278-287. doi: 10.19678/j.issn.1000-3428.0064802

• 开发研究与工程应用 • 上一篇    下一篇

基于增强型轻量深度网络的牧区牲畜高效检测

齐咏生1,2, 杜晓旭1, 朱俊峰1,3, 高胜利4, 刘利强1,2   

  1. 1. 内蒙古自治区机电控制重点实验室, 呼和浩特 010051
    2. 内蒙古工业大学 电力学院, 呼和浩特 010051
    3. 水利部牧区水利科学研究所, 呼和浩特 010051
    4. 内蒙古北方龙源风力发电有限责任公司, 呼和浩特 010050
  • 收稿日期:2022-05-25 出版日期:2023-07-15 发布日期:2023-07-14
  • 作者简介:

    齐咏生(1975—),男,教授、博士,主研方向为目标检测

    杜晓旭,硕士研究生

    朱俊峰,硕士

    高胜利,博士

    刘利强,教授、博士

  • 基金资助:
    国家自然科学基金(61763037); 内蒙古自治区机电控制重点实验室开放基金(IMMEC2020001); 内蒙古自治区科技计划项目(2021GG164); 内蒙古自治区自然科学基金(2021MS06018); 内蒙古自治区2022年度基本科研业务费项目(JY20220353)

Efficient Livestock Detection in Grazing Areas Based on Enhanced Lightweight Deep Network

Yongsheng QI1,2, Xiaoxu DU1, Junfeng ZHU1,3, Shengli GAO4, Liqiang LIU1,2   

  1. 1. Inner Mongolia Key Laboratory of Electrical and Mechanical Control, Hohhot 010051, China
    2. Institute of Electric Power, Inner Mongolia University of Technology, Hohhot 010051, China
    3. Pastoral Water Conservancy Science Research Institute of the Ministry of Water Resources, Hohhot 010051, China
    4. Inner Mongolia North Longyuan Wind Power Generation Co., Ltd., Hohhot 010050, China
  • Received:2022-05-25 Online:2023-07-15 Published:2023-07-14

摘要:

实现大数据管理牲畜需要实时监测牲畜,但对牲畜进行实时监测容易受到目标尺寸变化大、光照、环境因素等干扰,因此检测难度大,现有牲畜检测算法存在鲁棒性差等问题。提出一种基于增强型YOLOv4-tiny的目标检测算法(E-YOLOv4-tiny),采用多尺度特征融合的金字塔网络,兼顾浅层局部细节特征与深层语义信息,解决牧区牲畜尺寸波动问题。通过改进残差结构,减少主干网络参数量,以适应嵌入式平台需求。引入一种新的复合聚类算法设计锚框,在保证可移植性的前提下提高算法精度。针对牧区环境特点,提出一种新的复合多通道注意力机制,改善目标检测网络精度差的问题,增强算法鲁棒性。实验结果表明,E-YOLOv4-tiny算法的平均精度均值(mAP)为0.878 9,帧率为32帧/s,相较于传统YOLOv4-tiny算法,在保持几乎相同的检测速率条件下,mAP提升了9.32%。

关键词: 目标检测, 深度学习, 计算机视觉, YOLOv4-tiny算法, 注意力机制, 特征融合

Abstract:

Realizing big data to manage livestock requires real-time monitoring of livestock, but real-time monitoring of livestock is easily interfered by large changes in target size, lighting, environmental factors, etc., so it is difficult to detection, and existing livestock detection algorithms have the problem of poor robustness.An object detection network called E-YOLOv4-tiny is proposed based on enhanced YOLOv4-tiny, which adopts a pyramid network with multi-scale feature fusion, taking into account shallow local detail features and deep semantic information to solve the problem of livestock size fluctuation in pastoral areas.The number of backbone network parameters is reduced by improving the residual structure to accommodate embedded platform requirements.A new composite clustering algorithm is introduced to design anchor frames to improve the accuracy of the algorithm under the premise of ensuring portability.Finally, according to the characteristics of a pastoral environment, a new Compound Muti-channel Attention(CMA) mechanism is proposed to improve the poor accuracy of the target detection network and enhance the robustness of the algorithm.Experimental results show that the mean Average Precision(mAP) of the E-YOLOv4-tiny algorithm is 0.878 9, and the frame rate is 32 frame/s, and it's mAP is 9.32% higher than that of the traditional YOLOv4-tiny algorithm while maintaining almost the same detection rate.

Key words: target detection, deep learning, computer vision, YOLOv4-tiny algorithm, attention mechanism, feature fusion