An Improved Residual Network Based on Enhancing Projection Shortcuts Performance

Despite the popularity of the Tranformer architecture, convolutional neural networks still play an important role in computer vision. Among all convolutional neural network architectures, residual connected networks and Squeeze and Excitation blocks exhibit excellent performance. This article combines ResNet and SE locks, and only adds SE block in the projection shortcuts of shortcut connections. Our proposed method adds negligible parameters and computations than ResNet, and our model is simple than SE-Resnet. The experimental results on the CIFAR-10 dataset show that our method is superior to ResNet-110 and SE ResNet-110 at deeper layers of the network, and adding channel attention in projection shortcuts can greatly improve performance.