基于卷积神经网络的农机图像自动识别研究

Research on automatic recognition of agricultural machine image based on convolutional neural network

基于农机物联网技术的农机作业监管系统通过采集机具图像来判断农业机具类型和作业状态,但是由于图像数据量大,人工抽查方式工作量大、效率低,难以满足监管识别需求。构建包括播种机、翻转犁、起垄机、深松机和旋耕机等类型的图像数据集,并在Google公司的深度学习平台Tensorflow下对机具图像数据集进行标注和图像预处理。设计针对实际监管需求和图像特点的卷积神经网络模型,并通过减少过拟合与提高训练效率对模型进行优化。模型训练试验结果显示:本文设计的机具识别网络在验证集上的识别率达到98.5%,相同试验条件下,LeNet-5模型在验证集上的识别率为81%,ResNet-50模型在验证集上的识别率为98.8%,但是在识别效率上,ResNet-50模型完成训练需要近60 h,识别一张图片需要0.3 s,而本文设计的机具识别网络完成训练仅需要30 h,识别一张图片只需要0.1 s。为进一步验证模型的实用性,选取200张图像进行测试,测得模型对各类机具图像的精确度平均值为98.47%,召回率平均值为98.37%,F1-score平均值为98.41%,表明模型具有良好的鲁棒性,实用性良好。

The operation supervision system based on agricultural machinery networking technology can identify the machine type and the operation state by collecting the image of the machine tool.However,with increase in the amount of image data,manual sampling is faced with challenges such as having a heavy workload and little supervision,which does not meet the supervision requirements.In this paper,image data sets including seeder,tilting plough,erasing machine,deep looser and rotary cultivator were constructed,and the machine image data sets were annotated and preprocessed under Google’s deep learning platform Tensorflow.A convolutional neural network model was designed to meet the actual regulatory requirements and image characteristics,after which the model was optimized by reducing over-fitting and improving training efficiency.The model training experiment results showed that the recognition rate of the machine recognition network designed in this paper reached 98.5% on the verification set.Under similar experimental conditions,the recognition rate of LeNet-5 model and ResNet-50 model was 81% and 98.8%,respectively.However,in terms of recognition efficiency,ResNet-50 model needed nearly 60 hours to complete the training and 0.3 s to recognize a picture,while the machine recognition network designed in this paper needed 30 hours to complete the training,and 0.1 s to recognize a picture.In order to further verify the practicability of the model,200 images were selected for testing,and the average accuracy of the model for all kinds of machine and tool images was 98.47%,the average recall rate was 98.37%,and the average F1-score was 98.41%,indicating that the model had good robustness and practicability.