宿根蔗补种机轻量化蔗苗识别与定位技术

Lightweight Cane Seedling Identification and Positioning Technology for Root Cane Replanting Machine

摘要:

宿根蔗补种机是解决甘蔗田间缺苗问题的一种补种装置，对于补种机田间作业，能够精准地识别并定位宿根蔗苗十分重要。针对甘蔗田间宿根蔗苗难以精确检测并定位的问题，提出了一种双目相机结合改进YOLO v5目标检测算法的宿根蔗苗识别和定位方法。针对目标检测，提出了一种改进YOLO v5s网络模型YOLO v5s_P234_SGG。首先在不同光照及距离条件下拍摄宿根蔗苗图像，进行数据预处理和标注，构建宿根蔗苗数据集，然后剔除原始YOLO v5s网络模型的大目标检测层，新增一个小目标检测层，使模型能够更好地适应对蔗苗这种小目标的识别需求；其次在主干网络引入SimAM注意力机制，以增强模型对宿根蔗苗关键特征信息的关注，引入SlimNeck代替Neck网络，在保持足够精度的同时降低了模型复杂度，并将主干网络中的普通卷积模块替换成Ghost模块，显著减小了模型内存占用量。实验结果表明，该方法在宿根蔗苗数据集上精确率达到95.8%，召回率达到95.2%，平均精度均值达到97.1%，相比原始YOLO v5s网络，精确率上升3.1个百分点，召回率上升2.6个百分点，平均精度均值上升3.1个百分点，模型内存占用量减小7.7MB，参数量减少4062632，浮点运算次数减少7.8×109，单幅图像检测时间减少3.7ms。蔗苗定位实验结果表明，双目测距定位算法平均相对误差为0.97%，最大相对误差为4.60%。成功实现了对甘蔗苗的精准识别与测距，为后续的农业智能作业提供了重要的实时信息和决策支持。

Abstract:

The cane replanting machine is a replanting device developed to solve the problem of lack of seedlings in the sugarcane field, and it is very important for the field operation of the replanting machine to accurately identify and locate the cane seedlings. In order to solve the problem that it is difficult to accurately detect and locate cane seedlings in sugarseed field, a method for detecting and locating cane seedlings with binocular camera combined with improved YOLO v5 object detection algorithm was proposed. For object detection, an improved YOLO v5s network model was proposed YOLO v5s_P234_ SGG. Firstly, the pictures of cane seedlings were taken under different lighting and near and far conditions, and the data was preprocessed and annotated to construct the dataset of cane seedlings, and then the large target detection layer of the original YOLO v5s network model was eliminated, and a small object detection layer was added to make the model better adapt to the recognition needs of small targets such as cane seedlings. Secondly, the SimAM attention mechanism was introduced into the backbone network to enhance the model’s attention to the key feature information of the cane seedlings, and SlimNeck was introduced instead of the Neck network, which reduced the complexity of the model while maintaining sufficient accuracy and replacing the ordinary convolution module in the backbone network with the Ghost module, which significantly reduced the size of the model. Experimental results showed that the accuracy of the proposed method on the root cane seedling dataset reached 95.8%, the recall rate reached 95.2%, and the average accuracy reached 97.1%, compared with the original YOLO v5s network, the accuracy was increased by 3.1 percentage points, the recall rate was increased by 2.6 percentage points, the average accuracy was increased by 3.1 percentage points, the model volume was decreased by 7.7MB, the number of parameters were decreased by 4062632, the FLOPs was decreased by 7.8×109, and the detection time of a single image was decreased by 3.7ms. The results of the cane seedling positioning test showed that the average relative error of the binocular ranging and positioning algorithm was 0.97%, and the maximum relative error was 4.60%. The accurate identification and ranging of sugarcane seedlings were successfully realized, which provided important real-time information and decision-making support for subsequent agricultural intelligent operations.