【摘要】 针对大多数目标检测模型存在参数量较多、检测时间慢、缺乏空间定位能力的问题,提出1种基于轻量型LW-YOLOv3的棉花主茎生长点检测与定位方法。首先,对传统YOLOv3深度卷积神经网络架构进行改进,引入轻量级识别模型MobileNet实现对YOLOv3特征提取网络的替代,在保证检测精度的同时提高检测速度。其次,为了进一步降低参数量,用深度可分离卷积替换部分标准卷积,在不损失网络性能的前提下减少模型参数。最后,利用Realsense深度相机获取目标图像,建立了棉花主茎生长点三维空间坐标的计算模型。为验证基于LW-YOLOv3模型的棉花主茎生长点检测与定位方法的有效性,对获得的有限棉花主茎生长点图像数据集进行了样本扩充,并基于扩充后的数据集进行了LW-YOLOv3与其他轻量型模型的对比实验。结果表明,LWYOLOv3模型平均识别准确率可达到88.66%,平均每张图片检测时间为17.8ms,参数量与权重文件分别降低为原网络的10.85%和12.19%。棉花主茎生长点X、Y、Z三维空间定位平均误差分别小于3.04、 3.00、3.53mm。该方法在目标检测准确率与模型大小之间实现了平衡,有助于将来在机载低性能终端上实现对棉花主茎生长点检测与定位。更多还原

【Abstract】 Aiming at the problems of many parameters, slow detection time and lack of spatial positioning ability in most target detection models, a detection and positioning method of cotton main stem growth point based on lightweight lw-yolov3 was proposed. Firstly, the traditional architecture of YOLOv3 deep convolution neural network is improved, and the lightweight recognition model MobileNet is introduced to replace the YOLOv3 feature extraction network, which can ensure the detection accuracy and improve the detection speed. Secondly, in order to further reduce the number of parameters, the depth separable convolution is used to replace part of the standard convolution to reduce the model parameters without loss of network performance. Finally, the three-dimensional coordinate calculation model of cotton main stem growth point was established by using the Realsense depth camera to obtain the target image. In order to verify the effectiveness of the detection and location method of cotton main stem growth point based on LW-YOLOv3 model, the limited image data set of cotton main stem growth point was expanded, and the contrast experiment between LW-YOLOv3 and other lightweight models was carried out based on the expanded data set. The results show that the average recognition accuracy of LW-YOLOv3 model can reach 88.66%, the average detection time of each image is 17.8 ms, and the parameters and weight files are reduced to 10.85% and 12.19% of the original network respectively. The average positioning errors of X, Y, Z were less than 3.04 mm, 3.00 mm and 3.53 mm, respectively. This method achieves a balance between the accuracy of target detection and the size of the model, which is helpful for the detection and location of cotton main stem growth point on the airborne low-performance terminal in the future.更多还原