| 摘要: |
| 针对传统机器学习采用人工提取特征方法时，由于人为主观性而影响昆虫识别效果与计数准确性的问题，采用图像特征自动提取方法，将深度学习目标检测模型引入昆虫的识别与计数领域，对Faster-RCNN目标检测模型进行改进：针对昆虫体积小，图像分辨率较低的特点，用网络深度更深，运算量更小的深度残差网络（ResNet50）代替原来的VGG16，以提取更加丰富的特征；针对部分昆虫密集的特点，用Soft-NMS算法代替传统的非极大值抑制（NMS）算法，以减少密集区域的漏检。结果表明：改进后Faster-RCNN模型的检测准确率达到90.7%，较未改进的Faster-RCNN模型提高了4.2%，可以运用于昆虫的分类计数。利用深度学习目标检测模型进行昆虫识别与计数较传统的昆虫识别与计数方法更加方便，能够将昆虫的识别、定位和计数融为一体。 |
| 关键词: 昆虫识别 昆虫计数 Faster-RCNN 残差网络 Soft-NMS |
| DOI：10.11841/j.issn.1007-4333.2019.05.14 |
| 分类号: |
| 基金项目:徐州市重点研发项目（KC17055） |
|
| Insect identification and counting based on an improved Faster-RCNN model of the sticky board image |
|
ZHANG Yinsong, ZHAO Yindi, YUAN Muce
|
|
School of Environment and Surveying, China University of Mining and Technology, Xuzhou 221116, China
|
| Abstract: |
| In view of the traditional machine learning using artificial extraction feature method,due to the subjectivity of humans affecting the effect of insect recognition and the accuracy of counting,this study uses the automatic extraction of image features and introduces the deep learning target detection model into insect recognition.Compared with the counting field,the Faster-RCNN target detection model is improved:In view of the small insect size and low image resolution,the deep residual network (ResNet50) with deeper network depth and smaller computational capacity is used rather than the original VGG16.To extract more abundant features.Because of the high density of some insects,the Soft-NMS algorithm is used to replace the traditional non-maximum value suppression (NMS) algorithm in order to reduce the missed detection in dense areas.The results show that the improved Faster-RCNN model has an accuracy of 90.7%,which is 4.2% higher than that of the unmodified Faster-RCNN model,and can be applied to the classification and counting of insects.The use of deep learning target detection model for insect identification and counting is more convenient than traditional insect identification and counting methods.It can integrate insect identification,localization and counting. |
| Key words: insect identification insect count Faster-RCNN ResNet Soft-NMS |
