基于双线性卷积宽度网络的水稻病虫害识别

Rice pests and diseases recognition based on bilinear convolutional broad network

针对水稻病虫害数据集构建不够完善,现有方法对小样本数据集识别准确度低的问题,提出一种基于双线性卷积宽度网络(BCBN)的小样本水稻病虫害识别方法。所提方法利用双线性卷积神经网络(B-CNN)提取水稻病虫害图像的双线性特征,并通过宽度学习系统(BLS)算法增强双线性特征,从而提高模型的识别准确率。实验结果表明,BCBN在使用BLS增强双线性特征后,病虫害图像中的判别性特征得到了更高的权重占比,有效降低了模型的误分率,模型识别准确率达97.44%。所提方法在样本量较少时具有明显优势,能够满足真实场景下水稻病虫害分类检测的需求,为水稻病虫害识别技术提供了一种可行的思路。

Aiming at the problems that the construction of rice pest data set is not perfect and the existing methods have low recognition accuracy for small sample datasets,a small sample rice pests and diseases recognition method based on Bilinear Convolution Broad Network(BCBN)was proposed.A Bilinear Convolutional Neural Network(B-CNN)was used to extract the bilinear features of rice pests and diseases images,and the bilinear features were enhanced through the Broad Learning System(BLS)algorithm,thereby improving the recognition accuracy of the model.The experimental results show that BCBN enhances bilinear features by using BLS,and the discriminant features in pests and diseases images obtain a higher weight ratios,effectively reduce the error rate of the model,and the model recognition accuracy reaches 97.44%.The proposed method has obvious advantages when the sample size is small,which can meet the needs of classification and detection of rice pests and diseases in real scenes,and provides a feasible idea for rice diseases and insect pests identification technology.