Second-Order Response Transform Attention Network for Image Classification

Embedding second-order operations into deep convolutional neural networks (CNNs) has recently shown impressive performance for a number of vision tasks. Specifically, the two-branch second-order response transform (SoRT) network introduces the element-wise product transform into intermediate layers of CNNs, which facilitates the cross-branch response propagation and achieves promising classification accuracy. However, it fails to adaptively rescale responses of feature maps and largely changes the topology of the original backbone networks, leading to the limitation of generalizability. In order to overcome above problems, we propose a novel Second-order Response Transform Attention Network (SoRTA-Net) for classification tasks. The core of SoRTA-Net is the designed refined second-order response transform (RSoRT) module integrating reasonably the attention Squeeze-and-Excitation (SE) block and second-order response transform. Firstly, SoRTA-Net recalibrates adaptively feature responses by the SE block, and then the outputs are sequentially passed through the second-order response transform block, capturing approximately co-occurrence statistics and providing more nonlinearity. Finally, a shortcut branch is naturally combined with the output of the module to boost propagation. The proposed RSoRT module can be flexibly inserted into existing CNNs without any modification of network topology. Our SoRTA-Net extensively evaluated on three datasets (CIFAR-10, CIFAR-100, and SVHN). The experiments have shown that SoRTA-Net is superior to its baseline and achieves competitive performance.