A Combination Method of Stacked Convolutional Auto-Encoder and Selective Kernel Attention Mechanism for Image Classification

Convolutional auto-encoders have shown their remarkable performance in stacking deep convolutional neural networks for classifying image data during the past several years. However, they are unable to construct the state-of-the-art convolutional neural networks due to their intrinsic architectures. In this paper, we have proposed an effective stacked convolutional auto-encoder that integrates a selective kernel attention mechanism for image classification. This model is based on a fully convolutional auto-encoder and can be trained end-to-end. It consists of two parts: encoder and decoder. The encoder and decoder are composed of convolutional layer chain and deconvolution layer chain, respectively. The proposed method consists of three main modifications. First, a selection kernel (SK) convolution module and a selection kernel deconvolution module are constructed to form convolutional layer chain and deconvolution layer chain. Second, to solve the problem of network degradation, the idea of residual networks is cited to add skip layer connections between the SK convolution module and the symmetrically connected SK deconvolution module. Third, to alleviate the overfitting of the model, a method of adding noise in data augmentation is used to improve the generalization ability of the model. The experimental results show that this method can effectively integrate the channel attention module and the fully convolutional autoencoder. Although it is an unsupervised feature learning model, it can still achieve good classification results.