基于深度双线性转换注意力机制网络的林业有害生物识别方法

Forest Pest Identification Method Based on a Deep Bilinear Transformation Attention Mechanism Network

【目的】针对林业有害生物种类多,不少物种之间相似度高,视觉差异小,不易区分,导致林业防控人员无法快速准确识别有害生物种类的问题,本文提出基于深度双线性转换注意力机制网络(DBTANet)的林业有害生物细粒度图像识别方法。【方法】以自然状态下拍摄的60种林业害虫和14种林业有害植物图像作为研究对象,利用水平镜像、亮度调节、高斯模糊和高斯噪声等方法对图像数据集进行增强,按6∶2∶2比例划分为训练集、验证集和测试集;采用双线性插值法将每幅图像缩放至统一尺寸;改进ResNet网络中残差模块,加入深度双线性转换模块和注意力机制模块,建立DBTANet-101网络进行特征提取与分类;利用平均准确率、平均召回率和平均F1值3个指标评价不同模型对林业有害生物的识别结果。【结果】VGGNet-19、ResNet-50、ResNet-101、改进残差模块的ResNet-50和ResNet-101共5个模型对74种林业有害生物平均准确率分别为78.6%、74.9%、76.3%、79.7%和81.1%;在改进残差模块的ResNet-101基础上,增加深度双线性转换模块和注意力机制模块后,74种林业有害生物的平均准确率和召回率分别提高了10.2%和12.1%,22种相似的有害生物细粒度图像平均准确率提高了15.7%。【结论】基于深度双线性转换注意力机制网络(DBTANet)的林业有害生物细粒度图像识别方法,对74种林业有害生物和22种相似有害生物细粒度图像的平均准确率分别为91.3%和85.1%;双线性转换模块和注意力机制可以有效提高林业有害生物识别模型的准确率。

【Objective】There are many forest pest species,and some of them are morphologically similar with small visual differences,and are difficult to be distinguished,which makes it difficult for forestry prevention and control technicians to quickly and accurately identify them.In regard to those problems,this paper proposes a fine-grained image identification method of forest pests based on deep bilinear transformation attention mechanism network(DBTANet).【Method】The images of 60 forest pests and 14 harmful forest plants were targeted.The original images were enhanced by methods including horizontal mirror,brightness adjustment,Gaussian blur and Gaussian noise.All images were divided into training dataset,verification dataset and test dataset by a 6∶2∶2 ratio.The bilinear interpolation method was used to scale each image to a uniform size.The residual module in the ResNet network was improved.A deep bilinear conversion module and an attention mechanism module were added.DBTANet-101 model was developed to extract features and classify forest pest species.Three evaluation indexes of average accuracy,average recall rate and average F 1 score were used to evaluate the identification results of forestry pests by thd different models.【Result】A total of 74 forest pest species were identified by five models of VGGNet-19,ResNet-50,ResNet-101,improved ResNet-50 and improved ResNet-101,respectively,and the average accuracy rates of the five models were 78.6%,74.9%,76.3%,79.7%and 81.1%,respectively.On the basis of improved ResNet-101 of the residual module,the deep bilinear conversion module and the attention mechanism module were added,after then the average recognition rate and recall rate of 74 forestry pests increased by 10.2%and 12.1%,respectively.The average recognition rate of fine-grained images of 22 similar pests increased by 15.7%.【Conclusion】The fine-grained image recognition method of forest pests based on deep bilinear transformation attention mechanism network(DBTANet)achieves the average identification rate of 74 forest pests by 91.3%and the average identification rate of 22 similar forest pest species by 85.1%.The bilinear transformation module and the attention mechanism can effectively improve the accuracy of the forest pest identification model.