Multiple Feature Reweight DenseNet for Image Classification

Recent network research has demonstrated that the performance of convolutional neural networks can be improved by introducing a learning block that captures spatial correlations. In this paper, we propose a novel multiple feature reweight DenseNet (MFR-DenseNet) architecture. The MFR-DenseNet improves the representation power of the DenseNet by adaptively recalibrating the channel-wise feature responses and explicitly modeling the interdependencies between the features of different convolutional layers. First, in order to perform dynamic channel-wise feature recalibration, we construct the channel feature reweight DenseNet (CFR-DenseNet) by introducing the squeeze-and-excitation module (SEM) to DenseNet. Then, to model the interdependencies between the features of different convolutional layers, we propose the double squeeze-and-excitation module (DSEM) and construct the inter-layer feature reweight DenseNet (ILFR-DenseNet). In the last step, we designed the MFR-DenseNet by combining the CFR-DenseNet and the ILFR-DenseNet with an ensemble learning approach. Our experiments demonstrate the effectiveness of CFR-DenseNet, ILFR-DenseNet, and MFR-DenseNet. More importantly, the MFR-DenseNet drops the error rate on CIFAR-10 and CIFAR-100 by a large margin with significantly fewer parameters. Our 100-layer MFR-DenseNet (with 7.1M parameters) model achieves competitive results on CIFAR-10 and CIFAR-100 data sets, with test errors of 3.57% and 18.27% respectively, achieving a 4.5% relative improvement on CIFAR-10 and a 5.09% relative improvement on CIFAR-100 over the best result of DenseNet (with 27.2M parameters).