本文已被:浏览 2451次 下载 1508次 |
![本文二维码信息](/download_upload_file.aspx?file_name=/uploadfile/two_dimensional_code_img/zgnydxxb/2022/1/20220117.png) 码上扫一扫! |
基于轻量化YOLOv4的生猪目标检测算法 |
余秋冬1,2,3,杨明1,袁红3,梁坤3
|
|
(1.天津城建大学 计算机与信息工程学院, 天津 300384;2.天津职业技术师范大学 信息技术工程学院, 天津 300350;3.天津农学院 计算机与信息工程学院, 天津 300384) |
|
摘要: |
针对目前生猪目标检测算法模型较大,实时性差导致其难以在移动终端中应用等问题,将一种改进的轻量化YOLOv4算法用于生猪目标检测。在群养猪环境下以不同视角和不同遮挡程度拍摄生猪图像,建立生猪目标检测数据集。基于轻量化思想,在YOLOv4基础上缩减模型大小。结果表明,本研究算法的准确率和召回率分别为96.85%和91.75%,检测速度为62帧/s,相比于原模型,本研究算法在不损失精度的情况下,将模型大小压缩了80%,检测速度提高了11帧/s。本研究算法具有轻量化,稳健性强,实时性好的优点,能够更好地实现实际猪舍环境下生猪目标的检测,并有利于嵌入移动端设备中。 |
关键词: 生猪 深度学习 目标检测 轻量化 |
DOI:10.11841/j.issn.1007-4333.2022.01.17 |
投稿时间:2021-06-11 |
基金项目:国家级重点星火计划项目(2015GA610004);天津应用基础与前沿技术研究计划项目(16JCTPJC47000);天津市农业科技成果转化与推广项目(201404030) |
|
Pig object detection algorithm based on lightweight YOLOV4 |
YU Qiudong1,2,3,YANG Ming1,YUAN Hong3,LIANG Kun3
|
(1.School of Computer and Information Engineering, Tianjin Chengjian University, Tianjin 300384, China;2.School of Information Technology and Engineering, Tianjin University of Technology and Education, Tianjin 300350, China;3.School of Computer and Information Engineering, Tianjin Agricultural University, Tianjin 300384, China) |
Abstract: |
Aiming at the problem that the pig object detection algorithm is difficult to be applied in mobile terminals due to its large model and poor real-time performance, this study used an improved lightweight YOLOv4 algorithm on pigs object detection. In the actual pig house environment, pig images were taken under different perspectives and different occlusion degree in group rearing environment, and the dataset of live pig target detection was established. Based on the idea of lightweight, the model size is reduced on the basis of YOLOv4. The results showed that: The accuracy rate and recall rate of the proposed algorithm are 96. 85% and 91. 75%, respectively. The detection speed is 62 frames per second. Compared with the original model, the algorithm reduces the model size by 80% and improves the detection speed by eleven frames per second without loss of accuracy. The algorithm in this study has the advantages of light weight, strong robustness and good real-time performance, which can better realize the detection of pig object in the actual pig house environment and is conducive to be embedded in mobile terminal devices. |
Key words: pig deep learning object detection lightweight |