基于无人机图像和贝叶斯CSRNet模型的粘连云杉计数

自动、准确且快速地统计苗木数量是实现苗圃高效管理的重要基础。针对现有苗木计数方法准确率较低且无法准确统计粘连苗木等问题，该研究提出了一种基于贝叶斯CSRNet模型的云杉计数模型。该模型以对粘连苗木具有良好稳定性的CSRNet模型为基础，引入贝叶斯损失函数，以人工标注的点标签数据作为监督信号。以1 176幅云杉图像训练贝叶斯CSRNet模型，并通过166幅测试集云杉图像测试。结果表明，贝叶斯CSRNet模型可以准确、快速地统计无人机航拍图像内的云杉，对测试集图像内云杉的平均计数准确率（Mean Counting Accuracy，MCA）、平均绝对误差（Mean Absolute Error，MAE）和均方误差（Mean Square Error，MSE）分别为99.19%、1.42和2.80。单幅云杉图像耗时仅为248 ms，模型大小为62 Mb。对比YOLOv3模型、改进YOLOv3模型、CSRNet模型和贝叶斯CSRNet模型对166幅测试集云杉图像的计数结果，贝叶斯CSRNet模型的MCA分别比YOLOv3模型、改进YOLOv3模型、CSRNet模型高3.43%、1.44%和1.13%；贝叶斯CSRNet模型的MAE分别比YOLOv3模型、改进YOLOv3模型、CSRNet模型低6.8、2.9和1.67；贝叶斯CSRNet模型的MSE分别比YOLOv3模型、改进YOLOv3模型、CSRNet模型低101.74、23.48和8.57。在MCT和MS两项指标上，贝叶斯CSRNet模型与CSRNet模型相同且优于YOLOv3模型和改进YOLOv3模型。贝叶斯CSRNet模型可实现无人机航拍图像内苗木数量的自动、准确、快速统计，为苗木库存智能盘点提供参考。

Adhesion spruce counting based on UAV images and Bayesian CSRNet model

Abstract: An accurate and rapid counting of the seedlings is of great significance to evaluate the yield and quality in the nursery for the decision making on reasonable production plans. The traditional way is the manual operation to count the seedlings one by one by workers. It is time-consuming and labor-intensive, particularly with the slow data update. It is a high demand for the new counting of the seedlings to meet the actual needs. Taking spruce as the research object, this study aims to develop automatic, accurate, and fast counting, where the images were captured by the unmanned aerial vehicle (UAV) as the experimental data. At first, some attempts were made to count the seedlings using density map regression with deep learning, especially for the higher accuracy of the adhesion forest seedlings. Specifically, the density map regression (CSRNet) was selected as the base model, due to the excellent robustness of the seedling adhesion. In the CSRNet model, the coarse ground-truth density map was normally generated by the point annotations using a Gaussian kernel as the supervised signal training, leading to low counting accuracy. A Bayesian CSRNet model was designed to directly use the manually annotated point label data as the supervised signal, where the Bayesian loss function was also introduced to improve the counting accuracy of the CSRNet model. Then, the 558 acquired spruce images were divided into the training set (392 images) and test set (184 images), according to the ratio of 7:3. The brightness was also adjusted to add the random noise, in order to process 392 spruce images from the training set. Among them, the brightness adjustment mainly simulated the spruce images under different light intensities, whereas, the addition of random noise was the spruce images under different environmental noises. After data processing, the 392 spruce images in the training set were expanded to 1 176 images. After that, the Bayesian CSRNet model was trained on the 1 176 spruce images and tested on the 166 test set. The results show that the Bayesian CSRNet model can be used to accurately and quickly count the spruce in the UAV aerial images. The mean counting accuracy (MCA), mean absolute error (MAE), and mean square error (MSE) of the spruce counting in the test set images were 99.19%, 1.42, and 2.80, respectively. The Bayesian CSRNet model took only 248ms to count a single spruce image, and the model size was only 62 Mb after training. Finally, an experiment was designed to compare the counting of 166 spruce images in the test set by the YOLOv3, the improved YOLOv3, and the CSRNet model. It was found that the Bayesian CSRNet model presented an outstanding performance, in terms of MCA, MAE, and MSE. The MCA values of the Bayesian CSRNet model were 3.43%, 1.44%, and 1.13% higher than those of the YOLOv3 model, the improved YOLOv3 model, and the CSRNet model, respectively. The MAEs of the Bayesian CSRNet model were 6.8, 2.9, and 1.67 lower than those of the YOLOv3 model, the improved YOLOv3 model, and the CSRNet model, respectively. The MSEs of the Bayesian CSRNet model were 101.74, 23.48, and 8.57 lower than the YOLOv3 model, the improved YOLOv3 model and the CSRNet model, respectively. In terms of MCT and MS, the Bayesian CSRNet model was the same as the CSRNet model, but outperformed the YOLOv3 and improved the YOLOv3 model. The MCT of the Bayesian CSRNet model was 60 and 103 ms lower than the YOLOv3 and improved YOLOv3 model, respectively. The MS of the Bayesian CSRNet model was 173 and 206 Mb lower than the YOLOv3 and improved YOLOv3 model, respectively. Therefore, the Bayesian CSRNet model can be expected to serve as an effective way to realize automatic, accurate, and rapid statistics on the number of seedlings in UAV aerial images.