基于深度学习的近地面草原鼠洞识别计数关键问题研究与应用

Study on key problems for rat hole recognition and count near ground based on deep learning and its application

鼠洞密度可用于评估草原鼠害发生程度。在近地面鼠洞图片采集与识别中,图像分辨率和拍摄倾角是影响时效性和识别性能的关键因素。为此,设计了带有倾角传感器的图像采集装置,在锡林郭勒草原采集2325张鼠洞图片并进行手工标注,对比分析了3种图片输入尺寸(416 pixel×416 pixel、608 pixel×608 pixel、1024 pixel×1024 pixel)、4类拍摄倾角(21°、32°、41°、51°)、2种目标识别模型(YOLOv3和YOLOv4)对识别性能的影响。结果表明:YOLOv4模型在输入图片尺寸为416 pixel×416 pixel时能取得最优的性能。当拍摄倾角为41°时,识别性能最优;当拍摄倾角为32°时,识别性能最差。与近3年发表的鼠洞识别方法进行对比分析,验证了本文方法的性能先进性。研究结果可为草原鼠害的智能化监测提供技术支撑。

Rat hole density serves as a significant indicator for assessing the extent of rat damage in grasslands.In the image acquisition for rat hole recognition and count near ground by deep learning,the optimal image input size and shooting angle are crucial.To address these considerations,an image acquisition device equipped with an inclination sensor was designed,and a dataset of 2325 rat hole images was collected from Xilingol grassland and manually annotated.The recognition performance was compared under 3 image input sizes(416 pixel×416 pixel,608 pixel×608 pixel,1024 pixel×1024 pixel),4 shooting angles(21°,32°,41°,and 51°),and 2 target recognition models(YOLOv3 and YOLOv4).The findings revealed that YOLOv4 achieved superior performance when the image input size was set as 416 pixel×416 pixel.Furthermore,the recognition performance was optimal at the shooting angle of 41°,while it was the poorest at 32°.The proposed method was validated by comparing it with relevant approaches published within the past three years.These results offered valuable insights to support the development of intelligent monitoring technologies for assessing rat damage in grassland ecosystems.