卷积自编码器中粗粒度池化特征提取研究

卷积自编码器（Convolutional Auto Encoder，CAE）提取的粗粒度池化特征具有一定范围内旋转和平移的不变性，因而得到广泛使用.然而，目前CAE仍主要依靠经验调节内部参数以获取满足要求的粗粒度池化特征.本文将CAE看作一个整体，从概率上分析了影响其表现的具体原因，构建了一个通用框架用于调节其中的主要参数以获取更好的粗粒度特征.首先从概率上权衡了粗粒度特征在池化层上的判别性与不变性，并在CAE中选择合适的卷积范围和白化参数.然后通过分析池化域内特征的稀疏度选择相应的池化方法以获取具有更好可分离性的粗粒度池化特征.在两个公开数据库（STL-10和CIFAR-10）的实验结果表明本文提出的方法可以指导CAE提取到更好的粗粒度池化特征并在多类分类任务中表现得更好.

Coarse-Grained Pooled Features Learning in Convolutional Autoencoders

Coarse-grained pooled features obtained from convolutional autoencoder (CAE) achieve scale and shift invariances and have been widely used recently.However,in most previous works coarse-grained pooled features are obtained by empirically modulating parameters in CAE.In this paper,we see the CAE as a whole,find the probabilistic factors affecting the performance of it,and formulate a general framework to regulate parameters in it to obtain better coarse-grained representation.Firstly,the discrimination-invariance tradeoff of coarse-grained features is probabilistically evaluated in the pooled feature maps.Furthermore,the proper convolved filter scales and appropriate whitening parameters are suggested in a CAE.Secondly,pooling approaches are combined with the sparsity degree in pooling regions,and we propose the preferable pooling approach in different cases.Experimental results on two independent benchmark datasets (STL-10 and CIFAR-10) demonstrate that our framework can guide CAEs to extract better coarse-grained pooled features and performs better in multiclass classification task.