基于YOLOv4和双重回归的复杂环境檀香树缺苗定位方法

Missing Seedling Localization Method for Sandalwood Trees in Complex Environment Based on YOLOv4 and Double Regression Strategy

在檀香树大面积种植过程中,存在人工排查缺苗效率低、成本高和难以监管等问题,而且檀香树必备的伴生植物和树间穿插的其它作物,更加大了查补难度。针对这些问题,本文提出一种基于YOLOv4和双重回归的复杂环境檀香树缺苗检测和精准定位方法。首先,采用YOLOv4目标检测算法,处理无人机采集的遥感图像,实现檀香树植株的智能检测。然后,以双重线性回归结合延长列线补漏策略为核心,构建缺苗定位算法(Missing seedling localization algorithm, MSL):选任意檀香树作基准,根据像素坐标划分列区域,对各列区域中檀香树用线性回归拟合列线;对拟合后仍未归入列的遗漏檀香树,用延长回归线策略重新判断归属,并再次线性回归优化列线。最后,根据种植间距规划,实现缺苗检测和定位。试验结果表明,檀香树缺苗检测精确率86.82%、召回率82.25%、F1值84.47%、运行时间8.19 s。该方法融合了大疆无人机遥感图像采集系统的快速性、YOLOv4算法和双重回归策略的精准性,可实现对复杂生长状况下檀香树的实时智能缺苗检测和精准定位。

In the process of planting sandalwood trees on a large scale, there are problems such as low efficiency, high cost, and difficulty in the supervision of manual ranking of missing seedlings, and the necessary companion plants for each sandalwood tree and other crops interspersed between the trees, further deepening the difficulty of checking and replenishing. For these problems, a seedling deficiency detection and precise localization method in complex environment was proposed based on YOLOv4 algorithm and double regression strategy. Firstly, the YOLOv4 target detection model was used to achieve sandalwood plant detection from remote sensing images collected by UAV. Then the missing seedling localization algorithm(MSL) was constructed based on the double linear regression and extended column line fixing strategy: arbitrary sandalwood trees were selected as the benchmark, column regions were divided according to the pixel coordinates, and column lines were fitted to the sandalwood trees in each column region by using linear regression;for the omitted sandalwood trees that were not classified into columns after fitting, the attribution was judged again with the extended regression line strategy, and the column lines were optimized by linear regression again. Finally, the missing seedlings were calculated and localized according to the spacing at the time of planting. The results showed that the precision was 86.82%, the recall was 82.25%, the F1-score was 84.47%, and the running time was 8.19 s, respectively. In summary, this method combined the rapidity of DJI UAV remote sensing image acquisition system, the accuracy of YOLOv4 algorithm and double regression strategy, which can be used to achieve real-time intelligent seedling deficiency detection and accurate localization of sandalwood trees under complex growth conditions.