基于深度学习的诱捕器内红脂大小蠹检测模型

Detection Model of In-trap Red Turpentine Beetle Based on Deep Learning

红脂大小蠹是危害我国北方地区松杉类针叶树种的重大林业入侵害虫,其虫情监测是森林虫害防治的重要环节。传统的人工计数方法已经无法满足现代化红脂大小蠹监测的需求。为自动化识别并统计信息素诱捕器捕获的红脂大小蠹,在传统信息素诱捕器中集成摄像头,自动采集收集杯内图像,建立蠹虫数据集。使用K-means聚类算法优化Faster R-CNN深度学习目标检测模型的默认框,并使用GPU服务器端到端地训练该模型,实现了诱捕器内任意姿态红脂大小蠹的目标检测。采用面向个体的定量评价和面向诱捕器的定性评价两种评价方式。实验结果表明:较原始Faster R-CNN模型,该模型在困难测试集上面向个体和诱捕器的精确率-召回率曲线下面积(Area under the curve,AUC)提升了4. 33%和3. 28%。在整体测试集上个体和诱捕器AUC分别达0. 935 0、0. 972 2。该模型的检测速率为1. 6 s/幅,准确度优于SSD、Faster R-CNN等目标检测模型,对姿态变化、杂物干扰、酒精蒸发等有较好的鲁棒性。改进后的模型可从被诱芯吸引的6种小蠹科昆虫中区分出危害最大的红脂大小蠹,自动化地统计诱捕器内红脂大小蠹数量。

The red turpentine beetle (RTB) is a major forestry invasive insect that damages the coniferous species of pine trees in northern China. Therefore, the monitoring of RTB plays an important role in forestry pest controlling. However, traditional trap-based monitoring depends on human experts to manually recognize and count pests, which prohibits the modern RTB monitoring. To automatically recognize and count RTB captured by pheromone traps, a RGB camera was integrated in traditional cup trap to capture in-trap images and build the bark beetles dataset. The default boxes of Faster R-CNN object detection model based on deep learning were optimized by the K-means clustering algorithm. The optimized Faster R-CNN models were trained end to end by the GPU server, which enabled the in-trap detection of RTB with unconstrained postures. The models were evaluated by two metrics: the object oriented quantitative metric and the trap oriented qualitative metric. The experiments demonstrated that the optimized models outperformed the original Faster R-CNN model in terms of both metrics. The area under the curve (AUC) of precision-recall plot for object and trap on difficult test sets were increased by 4.33% and 3.28%, respectively. The AUC for object and trap on all test sets reached 0.935 0 and 0.972 2, respectively. The detection speed of the model was 1.6 s per image. The optimized models outperformed the SSD, Faster R-CNN object detection models in terms of accuracy, which was robust to pose variance, bark interference, alcohol volatilization, etc. The proposed method distinguished and counted RTBs from the six species of scolytidae insects attracted by the pheromone lure, which could reduce the human cost of pest monitoring and forecasting.