基于改进CenterNet的玉米雄蕊无人机遥感图像识别

为准确识别抽雄期玉米雄蕊实现监测玉米长势、植株计数和估产，基于无锚框的CenterNet目标检测模型，通过分析玉米雄蕊的尺寸分布，并在特征提取网络中添加位置坐标，从而提出一种改进的玉米雄蕊识别模型。针对雄蕊尺寸较小的特点，去除CenterNet网络中对图像尺度缩小的特征提取模块，在降低模型参数的同时，提高检测速度。在CenterNet特征提取模型中添加位置信息，提高定位精度，降低雄蕊漏检率。试验结果表明，与有锚框的YOLO v4、Faster R-CNN模型相比，改进的CenterNet雄蕊检测模型对无人机遥感影像的玉米雄蕊识别精度达到92.4%，分别高于Faster R-CNN和YOLO v4模型26.22、3.42个百分点；检测速度为36f/s，分别比Faster R-CNN和YOLO v4模型高32、23f/s。本文方法能够准确地检测无人机遥感图像中尺寸较小的玉米雄蕊，为玉米抽雄期的农情监测提供参考。

Improved CenterNet Based Maize Tassel Recognition for UAV Remote Sensing Image

In order to accurately identify the tassels of maize at tasseling stage, the growth, plant count and yield of maize should be monitored, based on the CenterNet object detection model without anchor frame, an improved maize tassel recognition model was proposed by analyzing the size distribution of maize tassels and adding position coordinates in the feature extraction network. According to the small tassel size, the feature extraction module for image scale reduction in CenterNet network was removed to reduce the model parameters and improve the detection speed. The location information was added to the CenterNet feature extraction model to improve the positioning accuracy and reduce the rate of tassel missed detection. The experimental results showed that, compared with YOLO v4 and Faster R-CNN with anchor frame, the improved CenterNet model achieved 92.4% accuracy in identifying maize tassels from UAV remote sensing images, which were 26.22 and 3.42 percentage points higher than that of Faster R-CNN and YOLO v4 models, respectively. The detection speed was 36f/s, 32f/s and 23f/s higher than that of the Faster R-CNN and YOLO v4 models, respectively. The method proposed can accurately detect the smaller tassels in the UAV remote sensing image, and provide a reference for the monitoring of agricultural situation in the tasseling stage of maize.