基于轻量二阶段检测模型的自然环境多类蔬菜幼苗识别

Identification of Multiple Vegetable Seedlings Based on Two-stage Lightweight Detection Model

摘要:

为实现自然环境下蔬菜幼苗精准快速识别，本文以豆角、花菜、白菜、茄子、辣椒、黄瓜等形态差异大、具有代表性的蔬菜幼苗为研究对象，提出一种基于轻量化二阶段检测模型的多类蔬菜幼苗检测方法。模型采用混合深度分离卷积作为前置基础网络对输入图像进行运算，以提高图像特征提取速度与效率；在此基础上，引入特征金字塔网络(Feature pyramid networks, FPN)单元融合特征提取网络不同层级特征信息，用于增强深度学习检测模型对多尺度目标的检测精度；然后，通过压缩检测头网络通道维数和全连接层数量，降低模型参数规模与计算复杂度；最后，将距离交并比（Distance-IoU, DIoU）损失作为目标边框回归损失函数，使预测框位置回归更加准确。试验结果表明，本文提出的深度学习推理模型对多类蔬菜幼苗的平均精度均值为97.47%，识别速度为19.07 f/s，模型占用存储空间为60 MB，对小目标作物以及叶片遮挡作物的平均精度均值达到88.55%，相比于Faster R-CNN、R-FCN模型具有良好的泛化性能和鲁棒性，可以为蔬菜田间农业智能装备精准作业所涉及的蔬菜幼苗检测识别问题提供新方案。

Abstract:

The identification of vegetable seedlings offers many useful applications in precision agriculture，such as automated weeding, variable rate fertilization and precise spraying of diseased plants. Aiming to recognize vegetable seedlings accurately and rapidly in natural environment, multiple kinds of vegetable seedlings were taken as study object, and a two-stage based lightweight detection model was proposed. In order to improve the speed and efficiency of image feature extraction, a mixed depth wise convolution that naturally mixed up multiple kernel sizes in a single convolution was applied as backbone network to process input images. Moreover, the feature pyramid networks (FPN) was introduced to integrate different feature maps of backbone network with the aim of improving the identification accuracy of deep learning detection model for multi-scale targets. By minimizing network channel dimensions and decreasing the number of full connection layers in detection head，the two-stage based detection model parameters and computational complexity were greatly reduced. In addition, a distance-IoU (DIoU) loss was proposed for bounding box regression to make the predicted box match with the target box perfectly. Experimental results showed that the mean average precision and recognition speed of multiple kinds of vegetable seedlings based on the proposed model were 97.47% and 19.07 f/s, respectively, and model size was 60 MB. The average accuracy of detection model can obtain 88.55%，when a crop size was less than 32 pixels×32 pixels or leaves occlusion occurred. It was demonstrated that the two-stage based lightweight detection model havd good generalization and robustness performance compared with that by other models，such as Faster R-CNN and R-FCN. The approach presented obtained high accurate rate and fast inference speed in the recognition of vegetable seedlings, which opened a new journey for the research of vegetable detection in precision agriculture.