基于改进YOLO v4网络的马铃薯自动育苗叶芽检测方法

为提高马铃薯幼苗叶芽检测识别的准确率,提高自动育苗生产系统的工作效率,提出了基于YOLO v4网络的改进识别网络。将YOLO v4特征提取部分CSPDarknet53中的残差块(Residual Block)替换为Res2Net,并采用深度可分离卷积操作减小计算量。由此,在增大卷积神经网络感受野的同时,能够获得叶芽更加细小的特征信息,减少马铃薯叶芽的漏检率。设计了基于扩张卷积的空间特征金字塔(D-SPP模块),并嵌入和替换到特征提取部分的3个特征层输出中,用于提高马铃薯叶芽目标识别定位的准确性。采用消融实验对改进策略的有效性进行了验证分析。实验结果表明,改进的识别网络对马铃薯叶芽检测的精确率为95.72%,召回率为94.91%,综合评价指标F1值为95%,平均精确率为96.03%。与Faster R-CNN、YOLO v3、YOLO v4网络相比,改进的识别网络具有更好的识别性能,从而可有效提高马铃薯自动育苗生产系统的工作效率。

Potato Leaf Bud Detection Method Based on Improved YOLO v4 Network

In order to improve the precision of the detection and recognition of the potato seedling leaf bud and improve the efficiency of the automatic seedling production system, an improved recognition network based on the YOLO v4 network was proposed. The residual block in the feature extraction part CSPDarknet53 was replaced with Res2Net, and the depthwise separable convolution was used to reduce the computation. In this way, the receptive field of the convolutional neural network can be enlarged, the finer feature information of leaf bud can be got, and the missed detection rate of potato leaf bud can be reduced. Furthermore, a spatial feature pyramid (D-SPP module) based on dilated convolution was designed and embedded in the output of the three feature layers of the feature extraction part to improve the recognition and localization precision of potato leaf bud target. The ablation experiment was used to verify the effectiveness of the improved strategies. The experiment results showed that the recognition precision, the recall rate, the comprehensive evaluation index F1 value and the average precision of the improved network were 95.72%, 94.91%, 95% and 96.03% respectively. Comparing with the common networks such as Faster R-CNN, YOLO v3 and YOLO v4, the improved network had the better recognition performances, thus the production efficiency automatic seedling production system can be enhanced.