基于残差网络与迁移学习的柑橘类型识别

传统的柑橘分类依靠人工进行辨识再手动完成分拣,整个过程耗费时间且成本高昂。对此提出了一种基于迁移学习与残差网络的柑橘图像分类方法。对Kaggle获取的20738张共8类柑橘的图像按7：3比例进行划分得到数据集。在此数据集上不同网络对于柑橘分类性能差异以及迁移学习对经典卷积模型在图像分类任务中的性能提升进行探究,实验以损失值、精准率、召回率等为性能评价指标。实验结果表明,在多种模型中,残差神经网络能获得比其他网络更高的准确率,使用迁移学习初始化网络参数能显著提高柑橘分类的准确度,降低模型过拟合的概率,实现对8类柑橘的准确识别分类,最终分类准确率达到99.9%,对柑橘分类具有指导意义。

Identification of citrus species based on residual network and transfer learning

Traditional citrus classification relies on manual recognition and then manual sorting, which is a time-consuming and costly process. A migration learning and residual network based citrus image classification method is proposed. A dataset of 20,738 images of 8 types of citrus obtained from Kaggle is divided into 7:3 ratio. The differences in performance of different networks for citrus classification and the performance improvement of migration learning on classical convolutional models in image classification tasks were investigated in this dataset, with loss value, precision rate and recall rate as performance evaluation metrics. The experimental results show that among the various models, the residual neural network can obtain higher accuracy than other networks, and the use of migration learning to initialise the network parameters can significantly improve the accuracy of citrus classification, reduce the probability of model overfitting, and achieve accurate recognition and classification of eight types of citrus, with a final classification accuracy of 99.9%, which is of guiding significance for citrus classification.