An effective image classification method for shallow densely connected convolution networks through squeezing and splitting techniques

The main drawbacks of traditional densely connected convolution networks (DenseNet) lie in: complex network models, excessive parameters, a large amount of computational and storage resources, falling into the problem of over-fitting, resulting in low object recognition accuracy. In addition, in the field of fine-grained image classification, the recognition performance is insufficient due to the inadequate representation capability of extracting features. In order to cope with these problems, we propose a novel shallow densely connected convolution networks (called DenseNet-S), it works as: (1) we adopt a shallow network training strategy to degrade the computational complexity and reduce the parameters, in order to avoid excessive number of layers affecting the recognition accuracy; (2) we propose a novel squeeze method to further reduce the network parameters and effectively alleviate the over-fitting phenomena. In addition, we apply the fire module and add the squeeze layer and the expand layer to the convolution module in DenseNet; (3) we employ the factorization technique into small convolutions, which can partition a large two-dimensional convolution into two small one-dimensional convolutions, in order to improve the feature extraction capability and the recognition performance in fine-grained image classification. The effectiveness of DenseNet-S was evaluated by extensive experiments on three benchmark datasets including CIFAR-10, CIFAR-100 and SVHN.