为了提高水稻虫害识别的准确率，文中首先采用深度语义分割 U-Net 网络去除复杂背景的影响，采用滑动窗口法提取水稻虫害图像的 HSV 颜色特征和 SILTP 纹理特征，统计同一水平滑窗中特征的最大值来构成特征向量，并利用 Relief-F 算法进行优化，获取具有高辨识性的水稻虫害图像特征。同时，引入对数据具有更好区分性的椭圆型度量，通过椭圆型度量学习寻找反映虫害图像特征空间结构信息和语义信息的非线性变换，对虫害图像特征的潜在关系进行建模，使相同类别特征之间的距离减小，不同类别特征之间的距离增大; 在椭圆型度量学习过程中，通过在三元组约束函数中增加 Fro-benius 范数正则项来避免过拟合，提高泛化能力。最后，利用椭圆型度量矩阵将水稻虫害特征变换到新的特征空间，从而提升 SVM 分类器的辨识能力。对 13 类常见水稻虫害图像的识别结果表明，文中提出的算法显著提高了小样本和复杂背景下水稻虫害图像识别的准确率，可以为精准农业中农作物病虫害的智能识别提供参考。

To improve the accuracy of identification of rice pest，an algorithm was proposed in this study． Firstly，the U-Net network of deep semantic segmentation was used to remove the influence of complex background． Then，the sliding window method was used to extract the HSV color feature and the scale-invariant local ternary pattern (SILTP) texture feature of rice pest images． The feature vector was formed by gathering the maximum value of the features in the same horizontal sliding window． Then the discriminative features of the rice pest images were ob-tained by the Relief-F optimization algorithm． At the same time，an elliptic metric with good distinguishability for data was introduced in this study． The nonlinear transformation that reflects the space structure and semantic infor-mation of rice pest image features was determined via elliptic metric learning． Then the model for the potential rela-tionship among rice pest image features was built． This model reduced the distance between same-class features but increases the distance between different-class features． In elliptic metric learning，the regular term of the Frobenius norm was added to the triple constraint function to avoid overfitting and improve generalization ability． Finally，the elliptic metric matrix was used to transform the features of the rice pest images into a new feature space． Thereby，the identification performance of the SVM classifier was improved． The identification results of 13 kinds of common rice pests show that the proposed method can significantly promote the few-shot identification accuracy of rice pest images in complex background and can provide valuable reference for the intelligent recognition of crop pest in pre-cision agriculture．