基于YOLOv5s-SPD的茶芽识别方法及识别系统光源设计与试验

Design and experiment of tea bud recognition method and light source of recognition system based on YOLOv5-SPD

实现茶芽的自动识别与定位是优质茶叶智能采摘设备研发的基础。针对茶芽细小,且采摘环境受光照影响较大等问题,本研究提出了一种基于深度学习网络模型的茶芽识别方法,开展识别系统的光源设计,能够为实现全天候和高效率的智能茶芽采摘设备提供技术支撑。首先,搭建铝合金框架的密闭遮光黑暗环境;然后,通过调节横杆高度和光源亮度创造出3种高度和3种光照强度组合;最后,采集不同组合情况下的茶芽图像数据集,利用改进YOLOv5模型对一芽一叶和一芽两叶开展识别测试。试验结果表明,YOLOv5s的总体准确率为77.13%,总体平均精度均值为86.14%,对于改进后的识别模型YOLOv5s-SPD的总体准确率为80.30%,总体平均精度均值为87.3%,单张图片的平均检测时间为5.7 ms,满足实时检测的要求,比原YOLOv5s总体准确率提升3.17%,总体平均精度均值提升1.16%,有效地提升了茶芽的识别性能。在高度90 cm和亮度L7(0.164~0.328μmol/m^(2))的条件下,一芽一叶和一芽两叶的检测准确率、召回率和AP平均值分别为86.70%、92.45%和95.00%。该方法可以有效快速地检测茶芽,光源设计方案为全天候优质茶叶智能采摘设备的研发提供了支持。

The realization of automatic identification and positioning of tea buds is the basis for the development of highquality tea intelligent picking equipment.Aiming at the problems of tiny tea buds and the picking environment which is greatly affected by light,this research proposes a tea bud recognition method based on a deep learning network model to carry out the light source design of the recognition system,which can provide technical support for the realization of all-weather and highefficiency intelligent tea bud picking equipment.First,an aluminum alloy frame was constructed to provide a closed and shaded dark environment;then,three combinations of heights and three combinations of light intensities were created by adjusting the height of the crossbar and the brightness of the light source;finally,the image datasets of tea buds in different combinations were collected,and recognition tests were carried out on one bud and one leaf and one bud and two leaves by using the improved YOLOv5 model.The experimental results show that the overall accuracy of YOLOv5s is 77.13%,and the overall average precision mean is 86.14%,the overall accuracy of the improved recognition model YOLOv5s-SPD is 80.30%,and the overall average precision mean is 87.3%,and the average detection time of a single image is 5.7 ms,which meets the requirement of real-time detection,and is better than the original YOLOv5s with an overall accuracy improvement of 3.17%and an overall average precision mean of 1.16%,which effectively improves the recognition performance of tea buds.Under the condition of height 90 cm and luminance L7(0.164~0.328μmol/m^(2)),the detection accuracy,recall and AP average of one bud and two leaves were 86.70%,92.45%and 95.00%,respectively.The method proposed in this paper can effectively and quickly detect tea buds,and the light source design scheme supports the development of all-weather high-quality tea intelligent picking equipment.