基于迁移学习的蔬菜图像识别方法

Vegetable image recognition based on transfer learning

为解决蔬菜识别领域缺少带标签样本的问题,提出了一种基于迁移学习的图像识别方法.首先,将原始数据集利用数据增强扩大样本数据量后引入到大规模数据集上的预训练模型.针对迁移过程中高层特征的领域特定性导致的网络泛化性能差,通过加入两层自适应层参数初始化后重新训练得到基本模型;对该基本模型再利用参数冻结的迁移方式进一步调优参数,得到用于蔬菜图像识别的最终网络模型.实验表明,基于CaffeNet和ResNet10两个小型网络的迁移策略可以较好地处理小样本的蔬菜图像识别,训练得到的模型准确率分别为94.97%、96.69%.与其他迁移算法及传统的神经网络方法相比,该算法具有更高的识别性以及更强的鲁棒性.

To solve the problem of lacking the labeled samples in vegetable recognition domain, a method for image recognition based transfer learning is proposed. Firstly, raw data are expanded by the data augmentation technique, then pre-trained models on the large-scale data sets are introduced to the target data set to train for the base model, where initializing two additional adaptive layers is performed. This step is aimed to solve the poor generalization performance caused by the transferred domain specific high-level features. Next, the final model is obtained for recognizing vegetable images by adopting the fine-tuning method with several layers frozen. Experimental results show that, based on CaffeNet and ResNet10, the proposed transfer approach reaches the accuracy of 94.97%, 96.69%, respectively, and can effectively process small samples in vegetable image recognition with higher accuracy and better robustness comparing to other transferring algorithms and the conventional convolution neural networks.