空间金字塔分解的深度可视化方法

发布时间：2018-04-21 03:38

本文选题：深度可视化 + 金字塔分解　；参考：《哈尔滨工业大学学报》2017年11期

【摘要】：针对基于深度卷积神经网络的图像分类模型的可解释性问题,通过评估模型特征空间的潜在可表示性,提出一种用于改善理解模型特征空间的可视化方法.给定任何已训练的深度卷积网络模型,所提出的方法在依据原输入图像使得模型类别得分激活最大化时,首先对反向传播的梯度进行归一化操作,然后采用带动量的随机梯度上升训练策略,反向回传修改原输入图像.引入了通过激活最大化获得的图像可解释性的正则化方法,常规正则化技术不能主动调整模型特征空间的潜在可表示性,结合现有正则化方法提出空间金字塔分解方法,利用构建多层拉普拉斯金字塔主动提升目标图像特征空间的低频分量,结合多层高斯金字塔调整其特征空间的高频分量得到较优可视化效果.通过限制可视化区域,提出利用类别显著性激活图技术加以压制上下文无关信息,可进一步改善可视化效果.对模型学习到的不同类别和卷积层中单独的神经元进行合成可视化实验,实验结果表明提出的方法在不同的深度模型和不同的可视化任务中均能取得较优的可视化效果.
[Abstract]:Aiming at the interpretable problem of image classification model based on deep convolution neural network, a visualization method is proposed to improve the understanding model feature space by evaluating the potential representability of the model feature space. Given any trained deep convolution network model, the proposed method first normalizes the backpropagation gradient when the model class score is maximized according to the original input image. Then the random gradient ascending training strategy of driving quantity is adopted, and the original input image is modified by reverse return. This paper introduces an interpretable regularization method of image which is obtained by maximizing the activation. The conventional regularization technique can not adjust the potential representability of the model's feature space actively. Combining with the existing regularization methods, the spatial pyramid decomposition method is proposed. The low frequency components of the target image feature space are actively promoted by constructing the multilayer Laplacian pyramid, and the high frequency component of the feature space is adjusted by the multi-layer Gao Si pyramid to obtain a better visualization effect. By limiting the visualization area, it is proposed that the context-independent information can be suppressed by using the category salient activation graph technique, which can further improve the visualization effect. The experimental results show that the proposed method can achieve better visualization results in different depth models and different visualization tasks.
【作者单位】：中国科学院成都计算机应用研究所;中国科学院大学;
【基金】：中国科学院西部之光人才培养计划项目
【分类号】：TP183;TP391.41

【相似文献】