实时监测蜜蜂觅食行为不仅有助于评估当地农作物授粉状况，也有助于及时采取措施提升蜜蜂产品的产量和品质。将计算机视觉技术引入蜜蜂觅食行为（是否携带花粉）的监测，设计5种浅层卷积神经网络（CNN）进行对比分析，并与深层网络GoogLeNet V1的迁移学习进行对比，寻找适合监测蜜蜂觅食行为的最优模型。发现GoogLeNet V1的识别准确率最高，达0.953 6，但训练耗时最长（7 326 s）；浅层卷积神经网络中准确率最高的模型为11层神经网络模型（含4层卷积层），测试准确率为0.903 6，耗时相对较短（1 054 s）。最后，对比传统机器学习算法，发现深度学习优势明显。研究表明，GoogLeNet V1深层网络适用于精度要求高、设备条件好的蜜蜂监测环境；而11层卷积神经网络更符合智能养蜂的实际需求。

Real-time monitoring of bee feeding behavior not only contributes to assess the pollination status of local crops,but also helps to take timely measures to improve the yield and quality of bee products.Computer vision technology was introduced into monitoring of the foraging behavior of bee(whether bee was carrying pollen or not).Five kinds of convolutional neural network(CNN)with differentdepths were designed and comparisons among them were made,and the deep network GoogLeNet V1 was compared with them,to find out the optimal model for monitoring the foraging behavior of bee.The results showed that,the GoogLeNet V1 had the highest recognition accuracy rate(0.953 6),whose training took up the longest time(7 326 s);The model with highest accuracy in shallow CNN was a network with 11 layers containing four convolutional layers,and its recognition accuracy was 0.903 6 with shorter training time (1 054 s).Finally, the results of comparison between CNN and traditional machine learning algorithms proved the distinct advantage of deep learning. In conclusion,the GoogLeNet V1 deep layer network is suitable for monitoring bee environments with high precision requirements and good equipment conditions.The 11-layer CNN designed is more in line with the actual needs of intelligent beekeeping.