摘要
针对YOLOv3模型定位目标边界框不够精确等问题,提出一种改进的YOLO算法。该算法在网络的残差模块中通过并行地引入通道注意力及空间注意力来提取关键目标特征,对多尺度预测与动物种类预测的方法做了新的设计。为证实算法效果,采集大量的原始图像数据,建立青藏高原地区牦牛、藏系羊和马等畜牧业动物图像数据集。数据集上新模型训练后,实验结果表明,算法提高了边界框的定位准确度和检测精度,相比YOLOv3算法,新算法在测试集上的mAP@IoU提高了1.6 mAP。
Aiming at the problem that the YOLO v3 model is not accurate enough to locate the target boundary box,an improved YOLO algorithm is proposed.The algorithm extracts key target features by introducing channel attention and spatial attention in the residual module of the network in parallel,and designs new methods for multi-scale prediction and animal species prediction.In order to verify the algorithm effect,a large number of original image data are collected,and image data sets of animal husbandry such as yak,Tibetan sheep and horse in Qinghai-Tibet Plateau are established.After training the new model on the data set,the experimental results show that the algorithm improves the positioning accuracy and detection accuracy of the bounding box.Compared with YOLO v3 algorithm,the mAP@IoU of the new algorithm on the test set is improved by 1.6 mAP.
作者
拉毛措
安见才让
拉毛杰
Lamao-cuo;Anjian-cairang;Lamao-jie(School of Information Science and Technology,Tibet University,Lhasa 850000,China;The Computer College of Qinghai Nationalities University,Xining 810007,China)
出处
《微处理机》
2021年第2期37-40,共4页
Microprocessors
基金
藏教财指[2018]81号
国家自然科学基金项目(61862054)。
关键词
畜牧业动物图像
目标检测
改进YOLO算法
通道注意力机制
空间域注意力机制
Animal images of stockbreeding
Target detection
Improved YOLO algorithm
Channel attention mechanism
Spatial domain attention mechanism
