A Co-Melting Reduction Auto-Encoder

The auto-encoder has been widely used in deep learning techniques as one of the unsupervised feature extraction algorithms. This paper proposes a Co-Melting Reduction Auto-Encoder (CMRAE) model to enrich extracted and release the burden of model parameters in image feature extraction. Firstly, a co-melting reduction structure is designed to realize feature fusion in the encoder, and it can reduce feature loss and model parameters in the decoder. Secondly, a multi-reconstruction loss function introduces reconstruction loss between feature layers, which can strengthen the connection between the feature layers and avoid the model premature effectively. The experiments prove that CMRAE can improve the feature extraction ability and efficiency of the model. The classification accuracy of features extracted by CMRAE is better than the state-of-the-art results on the datasets including Cats vs. Dogs (CvD), lung CT images, and multi-class weather (MWD).