【摘要】 针对物联网虫情测报灯对金银花害虫识别，存在光学环境条件难以满足导致识别效果差的问题，拟探索一种面向诱虫板的高效金银花害虫识别方法万物分割模型-残差网络（segment anything model residual network, SAM-ResNet），实现对3类金银花害虫（金龟类、棉铃虫、细蛾）进行计数和精准识别。首先运用万物分割模型（segment anything model, SAM）图像分割方法，基于图像编码器将诱虫板图像转换为低维特征向量，像素点作为提示信息转换为特征向量，两类特征向量共同传入掩码解码器，对害虫与害虫、害虫与背景板进行分割得到单一的害虫图像；对单一害虫图像进行标注重构为多分类数据集，运用Res Net对数据进行特征提取、前向传播、损失计算、反向传播以及参数优化等，训练模型不断优化。模型在虫情监测站提供的数据上进行试验，结果表明，在自然环境下SAM-Res Net可以无视光照、天气变化等影响，对诱虫板上害虫进行分割的准确率可达98.8%且耗时极短；在仅有600余张金银花害虫分割图像数据训练下，对3类害虫的识别准确率达到86.4%。此外，基于SAM-Res Net的设计原理，该方法理论上具有较高的普适性，可推广至其他农作物害虫的识别与监测领域，为虫害预警提供了技术支持。更多还原

【Abstract】 In view of the problem that the optical environment conditions are difficult to meet the requirements and the identification effect of honeysuckle pests is difficult to meet due to the insect detection and reporting lights of the Internet of Things insect detection and reporting lamp, it is proposed to explore a high-efficiency honeysuckle pest identification method(SAM-ResNet) for insect traps to achieve the counting and accurate identification of three types of honeysuckle pests(beetle, cotton bollworm and fine moth). Firstly, the Segment Anything Model(SAM) image segmentation method is used to convert the insect trap board image into lowdimensional feature vectors based on the image encoder, and the pixels are converted into feature vectors as prompt information. A single pest image was labeled and reconstructed into a multi-classification dataset, and the residual network ResNet was used to perform feature extraction, forward propagation, loss calculation,backpropagation and parameter optimization on the data, and the training model was continuously optimized.The results show that SAM-ResNet can segment pests on the insect trap board with an accuracy of 98.8% and a very short time in the natural environment, regardless of the influence of light and weather changes. Under the training of only more than 600 honeysuckle pest segmentation image data, the identification accuracy of the three types of pests reached 86.4%. The SAM-ResNet method can realize the accurate segmentation and counting of insect trap pests, and can accurately identify the pest species, which provides technical support for the early warning of honeysuckle insect pests.更多还原