基于Mixup算法和卷积神经网络的柑橘黄龙病果实识别研究

目的 解决传统柑橘黄龙病果实图像识别方法依赖人工设计特征、费时费力、网络模型参数量大、识别准确率低等问题。方法 首先，采集柑橘黄龙病的果实图像，并对其进行翻转、旋转、仿射、高斯扰动等数据扩增；采用Mixup算法建立样本之间的线性关系，增强模型识别数据样本的鲁棒性；然后，迁移Xception网络在ImageNet数据集上的先验知识，提出一种基于Mixup算法和卷积神经网络的柑橘黄龙病果实识别模型-X-ResNeXt模型；最后，采用动量梯度下降优化方法，有效地减缓震荡影响，并且有效地加速模型向局部最优点收敛。结果 采用数据扩增数据集训练的X-ResNeXt模型准确率可以达到91.38%；在进行迁移学习优化后，训练时间减少了432 s，准确率提升为91.97%；结合Mixup混类数据增强进一步训练，模型准确率提升为93.74%；最后，利用动量梯度下降方法进行模型收敛优化，最终模型的准确率达到94.29%，比Inception-V3和Xception网络分别提高了3.98%和1.51%。结论 在数据量较少情况下，降低模型复杂度并迁移已有先验知识，有助于模型性能提升；Mixup混类数据增强方法有利于提高模型识别柑橘黄龙病果实图像样本的适应性，提升柑橘黄龙病果实识别模型性能；X-ResNeXt模型在准确率与召回率指标上优于经典识别模型，可为柑橘黄龙病的高精度、快速无损识别提供参考。

Identification of citrus fruit infected with Huanglongbing based on Mixup algorithm and convolutional neural network

Objective The traditional image recognition method relies on manual design features, is time-consuming and labor-intensive, has large number of network model parameter and has low recognition accuracy rate. The goal was to solve these problems in traditional method for identifying citrus fruit infected with Huanglongbing.Method Firstly, we collected the images of citrus fruit with Huanglongbing, and performed data enhancement modes such as flip, rotation, affine, and Gaussian disturbance. Further, we used the Mixup algorithm to establish a linear relationship between samples to enhance the robustness of the model for identifying data samples. Then, we transfered the prior knowledge on the ImageNet data set of Xception network, and proposed a citrus Huanglongbing fruit recognition model of X-ResNeXt model based on Mixup algorithm and convolutional neural network. Finally, the momentum gradient descent optimization method was used to reduce the impact of shocks and effectively accelerate the convergence of the model to the local optimum.Result The accuracy rate of the X-ResNeXt model trained on the data set after data enhancement was 91.38%. After optimization with transfer learning, the training time reduced by 432 s, and the accuracy rate of the model increased to 91.97%. Combined with the enhancement of Mixup mixed data for further training, the accuracy rate of the model improved to 93.74%. Finally, the momentum gradient descent method was used to optimize the model convergence, and the final model accuracy rate reached 94.29%, which was 3.98% and 1.51% higher than Inception-V3 and Xception networks, respectively.Conclusion In the case of a small amount of data, reducing the complexity of the model and transfering existing prior knowledge will help to improve the performance of the model. The Mixup mixed data enhancement method is beneficial to improve the adaptability of the model to identify image samples of citrus fruit with Huanglongbing and improve model performance. The X-ResNeXt model is superior to the classic recognition model in terms of accuracy rate and recall rate, and can provide references for the high-precision, rapid and non-destructive identification of citrus Huanglongbing.