基于循环卷积神经网络的目标检测与分类

摘要

卷积神经网络模仿人类的视觉识别能力,提取图像目标的显著抽象特征,在图像目标检测与分类的应用上效果良好.在当前比较流行的批量随机梯度训练算法训练卷积神经网络的过程中,当神经元处于饱和状态时,会出现梯度下降缓慢和过度拟合问题,易使神经网络模型训练陷入困难.结合卷积神经网络和循环神经网络的特点,提出了构造浅层循环卷积神经网络,且在训练循环卷积神经网络模型时,分别采用进退法、黄金分割法自适应地改变批量随机梯度下降算法的规范化参数和学习率.实验结果表明,改进算法能够较好地避免梯度下降缓慢和过拟合问题,在训练循环卷积神经网络模型时具有较好的目标检测分类效果和更快的收敛性.%Convolutional neural network simulates human vision recognition and extracts the abstract characteristics significantly of the image target,with better effects on the application of image target detection and classification.In the currently popular training of convolution neural network by batch stochastic gradient algorithm,when neurons in a saturated state,there will be a slow gradient descent and excessive fitting which lead to the difficulties in training of the neural network model.In this paper,we propose a simple circular convolutional neural net-works combined with the characteristics of the convolutional and circular neural networks.In the training of circular convolutional neural net-work model,advance and retreat method and golden section are used to adaptively change the normalized parameter and the learning rate of batch stochastic gradient descent algorithm.Experiment shows that the proposed algorithm,with a better effect on detection and classification with faster convergence,can avoid the problem of slow gradient decent and excessive fitting to some extent.