在反卷积网络中引入数值解可视化卷积神经网络

Introduce Numerical Solution to Visualize Convolutional Neuron Networks Based on Numerical Solution

经典的反卷积可视化模型通过反池化、反激活、反卷积将特征图像还原至原图像空间,可视化网络节点从输入图像学习到的特征,有助于探究卷积神经网络运行良好的机制,但是由于采用近似处理,还原特征不明显。本研究引入数值求解方法来代替原模型中直接用卷积核的反转近似反卷积核的方法。先构造数据集:随机生成大小、形状、位置不一的结构简单、角点特征明显的三角形和矩形,用于组成层次结构逐渐复杂的数据集,并利用该数据集测试模型的可视化效果。实验表明,改进后的可视化模型提取的特征更明显,引入的噪音更少,可以更为精确地将激活网络节点从原图像学习的特征可视化。在更大的数据库上进行实验来验证结果,并利用这种结果进一步探究准确率与网络结构之间存在何种关系。

Zeiler's visualization model restore the feature maps to original image space,by unpooling and deconvolution,to visualize what the node learn from the image.It helps to research the convolutional neural networks mechanism,but the result is not apparent for the vague method.Based on the Zeiler's deconvolutional visualiztion model,numerical solution method was introduced to replace the vague method that just use convolutional kernel.The database was constructed firstly.The triangle and rectangle was generated with random size,shape and location,which have simple structure and apparent vertex.Based on the database,we constructed hierarchy database and took out experiment.The experiment results show that the improvement model extracts more apparent features and has less noise,which has more precise result.Experiment on bigger database was taken to verify our result,and the result to guide how to construct the network's stucture.