Abstract: To address the challenges of difficult identification, low efficiency, and low accuracy in grassland locust prevention and control surveys, an intelligent locust species identification APP system based on an improved ResNet50 model was proposed. The objective of this study was to utilize a dataset of 4454 photographs of various locust species, captured under diverse environmental conditions by using mobile devices. The training outcomes of six classification models were comparedby using the Adam optimizer and a cosine annealing learning rate schedule as GoogLeNet, AlexNet, VGGNet16, ResNet34, ResNet50 and MobilenetV3, with the aim of selecting the most optimal network. Then, the models were further refined by introducing an attention mechanism, which led to an improvement in their predictive accuracy. With the enhanced network, backend integration and frontend design were carried out, culminating in the development of a locust recognition application. Experimental results indicated that the average accuracy of the improved model was increased to 98.9%. The accuracy of the test set was improved to 96.6%, which represented a 7% increase compared to before the enhancement. This system which could be installed on mobile devices to ensure that detailed information regarding locust occurrences was accurately captured during surveys.

Key words: grassland locusts, monitoring, deep learning, attention mechanism, improved ResNet50

摘要： 为解决草原蝗虫在防控调查中识别困难、时效低下、准确率低等问题，提出基于改进ResNet50的蝗虫种类智能识别APP系统。以移动端设备在不同环境下拍摄的4454张不同种类的蝗虫图片为基础，采用Adam优化器与余弦退火的学习率退火方式于GoogLeNet、ALexNet、VGGNet16、ResNet34、ResNet50、MobilenetV3六种分类模型训练成果对比下，挑选最优网络。加入注意力机制，提升模型准确率，又以改进后的网络为识别模型对其进行后续的端口接入与前端的平面设计，最终形成蝗虫识别APP。试验表明：改进后的模型平均准确率提升至98.9%；测试集中的准确率为96.6%，比改进前提高7%。该蝗虫识别APP系统可安装至移动端设备，以确保蝗虫调查时准确把握蝗虫发生的详细信息。

关键词: 草原蝗虫, 监测, 深度学习, 注意力机制, 改进ResNet50