王宏乐1,2,3,王兴林 2,李文波 3,邹阿配 4,叶全洲 4,刘大存 4.一种基于解耦旋转锚框匹配策略的谷粒检测方法[J].广东农业科学,2022,49(12):144-150 |
查看全文
HTML
一种基于解耦旋转锚框匹配策略的谷粒检测方法 |
A Rice Grain Detection Method Based on Rotation-decoupled Detector for Oriented Object |
|
DOI:10.16768/j.issn.1004-874X.2022.12.016 |
中文关键词: 水稻 谷粒检测 旋转框 长宽比 目标检测 计算机视觉 |
英文关键词: rice grain detection oriented bounding box aspect ratio object detection computer vision |
基金项目:深圳市科技计划项目(CJGJZD20210408092401004) |
|
摘要点击次数: 1025 |
全文下载次数: 736 |
中文摘要: |
【目的】基于计算机视觉技术建立水稻考种中谷粒计数与粒径的检测方法,节省考种人力成本,提升效率,有利于数字化育种体系建立。【方法】通过对比基于 YOLOv5l 骨架结构的水平框目标检测策略与基于 ResNet101 和 DarkNet53 骨架结构的旋转框检测策略,同时比较旋转框检测策略预测的水稻谷粒长宽比值与实际测量值之间的差异,验证该方法的可行性。【结果】旋转框的目标检测模型计算得到的水稻谷粒长宽比与实测值无显著差异,两者的均方根误差为 0.91~2.29;而水平框的目标检测模型计算的长宽比显著小于实测值,且均方根误差值(9.38~9.45)比旋转框的目标检测模型更大。基于水平框的目标检测策略与基于旋转框的检测策略在谷粒计数的精度几乎相当,但水平框检测模型无法实现水稻谷粒长宽比的准确计算,而使用旋转框的目标检测策略能够较为准确计算谷粒长宽比。【结论】基于解耦旋转锚框匹配策略可对水稻谷粒进行准确计数,相比传统水平框的目标检测策略,可降低检测中的背景噪声和密集堆积物体的漏检情况,同时快速准确计算水稻谷粒长宽比。该方法可以进一步应用到水稻品种识别中的谷粒长短和大小的计算,以及种子质量检测及品种鉴定等数字化育种场景中。 |
英文摘要: |
【Objective】In rice seed investigations, the detection methods for rice grain counting and aspect ratios estimating were established based on computer vision technology, which could save labor costs and improve efficiency. It is conductive to the establishment of digital breeding system.【Method】The feasibility of rice detection methods was verified through comparing the horizontal bounding boxes (HBB) detection strategy based on model of YOLOv5l and oriented bounding boxes (OBB) detection strategy based on models of ResNet101 and DarkNet53. Meanwhile, the differences between the rice grain aspect ratios predicted by the OBB detection strategy and the actual measured values were compared. 【Result】No significant difference was found between the aspect ratios of rice grains calculated by the OBB detection models and the measured values, and both the values of root mean square error (RMSE) ranged from 0.92 to 2.29. The aspect ratios calculated by HBB detection models were significantly lower than the measured values, with greater values of RMSE ranging from 9.38 to 9.45. The accuracies of OBB detection strategy are almost equivalent to those of HBB detection strategy in rice grain counting. The OBB detectors could accurately calculate the aspect ratios of rice grains, while the HBB detectors could not.【Conclusion】The study results demonstrated that the OBB detection could count rice grains accurately. Compared with the traditional target detection strategy of HBB, it could reduce the background noise, lower the inaccuracy of densely distributed
objects, and calculate the aspect ratios of rice grains accurately. The method used in the study could be further applied to calculation of length, width and size of rice grains, detection of seeds quality, identification of seed cultivars and other digital breeding system. |
查看/发表评论 下载PDF阅读器 |