自适应调节学习率和样本训练方式的场景分类

Scene Classification with Adaptive Learning Rate and Sample Training Mode

在基于卷积神经网络的场景分类方法中,当训练数据较少时,由于网络训练次数较多、收敛性较差等原因,容易造成过拟合.为了消除此影响,在卷积神经网络的框架下,文中提出可以自适应调节网络学习率和样本训练方式的场景分类算法.根据网络训练中误差函数的变化自适应调节学习率,当误差函数变化较小时,保持批次的学习率不变,当误差函数变化加大时,学习率的变化与误差函数的改变成反比.同时根据网络输出结果改变实验样本的训练方式,着重训练分类不准确的图像.在Scene-15、Cifar-10场景数据集上的实验表明,文中算法改善神经网络的收敛性,有效提高分类精度,特别是对于室内场景等特征复杂场景的分类精度.

In scene classification based on convolutional neural network, over-fitting is caused due to the large number of network training and poor convergence performance with the small training dataset. To eliminate the negative effect, an algorithm for scene classification with adaptive learning rate and sample training mode is proposed. The network learning rate is adaptively adjusted on the framework of convolutional neural network according to the variation of the error function in the network training. When the error function changes slightly, the learning rate of the batch is unchanged. When the error function changes more remarkably, the variation of the learning rate is inversely proportional to the variation of the error function. Meanwhile, according to the network output, the sample training mode is switched, and the inaccurately recognized images are emphatically trained. The experimental results on Scene-15 and Cifar-10 scene datasets show that the proposed method improves the convergence of neural networks and effectively improves the classification accuracy, especially the classification accuracy of complex scenes such as indoor scenes.