基于卷积网络特征迁移的小样本物体图像识别

针对小规模物体图像识别过程中训练样本不足、准确率低,而深度卷积神经网络模型开发难度较大等问题,提出一种基于特征迁移的方法。首先对数据进行归一化等预处理,然后从预训练模型迁移不同层的特征到小样本数据集的训练中,修改预训练模型的分类层参数。最后以验证集的正确率和损失率作为评估指标,在GRAZ-02数据集上进行实验。结果表明,通过迁移VGG16网络的底层特征,再训练顶层网络,图像识别的正确率可达到95.28%,相比普通卷积神经网络模型可提高20个百分点。通过实验证实了采用特征迁移方法可以对小样本物体图像数据集开发出性能良好的深度神经网络图像识别模型。

Small SampleObject Image Recognition Based on Convolutional Network Features Transfer

Aiming at the problem of insufficient training samples and low accuracy in the object recognition process of small scale image data sets, and the difficulty of developing deep convolution neural network models, a new method based on network feature migration is proposed. First, the data were normalized, and the characteristics of different layers were migrated from the pre training model to the small training sample data set, then classification layer parameter of the pre training model was modified. Finally, taking the correct rate and loss rate of the validation set as the evaluation index, the experiments were carried out on the GRAZ-02 data sets. The results show that by migrating the underlying features of VGG16 network and training the top level network, the accuracy of image recognition can reach 95.28%, which is 20 percentage points higher than that of general convolutional neural network. The experiments prove that the feature migration method can be used to develop a deep neural network image recognition model with good performance for small sample object image data sets.