基于样本倍增、深度神经网络与SVM的少样本图像识别技术

Image recognition technology for few samples based on sample multiplication,deep neural network and SVM

基于深度神经网络的图像识别技术在少样本情况下会出现识别准确率下降,甚至训练失败的情况。本文针对该问题,提出了一种将少样本倍增、深度神经网络、SVM(support vector machine)相结合的算法。通过样本倍增提升样本数量,通过深度神经网络训练得到图像的高维特征,通过经典特征提取算法提取图像常规特征,构建包含高维特征和常规特征的支持向量,使用SVM进行训练。实验表明,本文所提方法对于特定样本集,在样本数量为总样本1/40的情况下,识别准确率达到61.36%(仅仅使用YOLO神经网络算法识别准确率为44.08%),大大提高了少样本下目标识别准确率,具有重要的应用价值。

In the case of few samples,the recognition accuracy of image recognition technology based on deep neural network will decrease.To solve this problem,this paper proposes an algorithm which combines sample multiplication,deep neural network and SVM(support vector machine).Firstly,the sample multiplication is used to increase the number of samples.Secondly,deep neural network is used to obtain the high-dimensional features of the image.Thirdly,the classical feature extraction algorithms is performed to extract conventional features of the image,and then the support vector including the high-dimensional features and the conventional features is constructed,which is trained by SVM.Experiments show that the accuracy of the proposed method is 61.36%(the accuracy of using Yolo algorithm only is 44.08%)for a specific sample set,whose number of samples is 1/40 of the total samples,which greatly improves the accuracy of recognition under a small number of samples and has important application value.