基于多尺寸特征图卷积方法的玉米雄穗检测

为了解决传统雄穗检测方法因玉米品种不同以及田间环境不同导致的检测误差较大、鲁棒性较差的问题,利用深度卷积神经网络提取特征,并对多尺寸特征图卷积的方法检测玉米雄穗。采用深度卷积神经网络inception作为基础网络来训练提取玉米雄穗特征,同时增加额外的卷积层对图像进行卷积提取特征,最后分别对基础网络中的两层卷积层以及额外的卷积层卷积得到的不同尺度特征图进行分类和位置回归。整体网络结构是多尺度端到端框架,效率高,方便检测不同尺度的雄穗。实验结果表明,此方法提高了雄穗检测的速度和准确率。

A Maize Tassel Detection Based on Method of Multi-Scale Feature Map Convolution

In order to solve the problem that traditional tassel detection methods based on image processing have large detection errors and poor robustness caused by different maize varieties and different field conditions. This paper proposes the use of a deep convolutional neural network to extract features, and detects corn tassels using a multi-scale feature map convolution method. The deep convolution neural network inception as a basic network was used to train and extract the feature of maize tassel. At the same time, the additional convolution layers were added to perform convolution and extraction on the images. Finally, classification and position regression were performed on the different scales feature maps calculated from the two convolution layers in basic network and the additional convolution layers. The overall network structure of this method is a multi-scale and end-to-end framework, which has high efficiency and is convenient to detect tassels at different scales. Experimental results show that this method improves tassel detection speed and accuracy.