基于无人机可见光影像和点目标检测的烟草烟株计数

Counting Tobacco Plants Based on UAV Visible Images and Point-Wise Object Detection

摘要: 无人机是一种灵活、便携、实时、高效的影像采集工具，被普遍应用于烟草烟株定位及计数，本研究以烟草烟株为研究对象，以无人机影像为数据，提出一种新的烟草烟株记数深度学习模型，该模型区别于传统的基于外围矩形框的目标检测，本文以中心点预测为目标学习烟草区域尺度形态，并采用轻量级的编、解码器从无人机遥感影像快速识别烟草。首先，本文提出的模型针对烟草植物形态学特点，通过中心关键点标注的方法，采用了基于SSD (Single Shot Multibox Detector)的多层特征融合方法，将来自不同深度层次的特征图进行融合，有效提高了目标检测的准确率。其次，对比分析了检测模型在不同高度的图像下的检测精度，本文提出的CDNet平均检测精度 > 98.89%，满足业务化应用的需求。本文提出的烟草烟株计数深度学习模型能够准确地检测不同飞行高度和不同生长期的无人机遥感影像中的烟草烟株，为烟草烟株的生长监测提供可靠数据支持。

Abstract: Drones are flexible, portable, real-time, and efficient tools for locating and counting tobacco plants. This study proposes a new deep learning model that detects tobacco plants from drone remote sensing images. Unlike traditional object detection methods that use rectangular boxes to enclose the targets, this model predicts the center point of each tobacco plant and learns its scale and shape. The model also uses a lightweight encoder and decoder to quickly identify tobacco plants from the images. The main contributions of this paper are as follows: First, the model adapts to the morphological characteristics of tobacco plants and uses a center key point annotation method. It also employs a multi-layer feature fusion method based on SSD (Single Shot Multibox Detector) to combine feature maps from different depth levels, which effectively improves the detection accuracy. Second, the model is tested on images at different heights and compared with other detection methods. The average detection accuracy of the proposed CDNet is higher than 98.89%, which meets the requirements of practical applications. The proposed deep learning model can accurately detect tobacco plants in drone remote sensing images with varying flight altitudes, growth stages, and resolutions, providing reliable data support for monitoring the growth of tobacco plants.