SReResNet: A stage recursive residual network for suppressing semantic redundancy during feature extraction

Feature extraction neural networks are essential components of computer vision systems. As the most famous one, ResNet has been widely used in engineering. Although many modern networks have outperformed ResNet, the costly pretraining and hyper-parameter optimization processes prevent their application of them in industrial computer vision systems. To avoid these costly processes, a Stage Recursive Residual Network (SReResNet) is proposed, which merely adjusts the forward propagation pipeline without changing the parameter architecture of ResNet. Thus, it can inherit existing model parameters of ResNet and replace ResNet through simple fine-tuning. Moreover, SReResNet is the first network to improve accuracy by utilizing the semantic trend during feature extraction instead of well-designed modules. It models a series of cascaded modules as a semantic unit and feeds the high-level feature maps back to the low-level modules for further semantic redundancy suppression through one feedback connection within the unit based on the semantic trend and the mechanism of looking and thinking twice. For object detection tasks, a Stage Recursive Feature Pyramid Network (SReFPN) is also proposed to rethink and suppress semantic redundancy further. Experiments demonstrate that SReResNet outperforms its counterparts in object detection and image classification tasks. On the MS COCO 2017 dataset, SReResNet outperforms ResNet with 1.2 Box AP improvement, and SReResNet with SReFPN further achieves 2.5 Box AP improvement without bells and whistles. On the CIFAR-100 dataset, SReResNet outperforms its counterparts, such as ResNet, DenseNet, and ConvNeXt, with at least 2.33% top-1 acc improvement. The code is available at https://github.com/unbelieboomboom/SReResNet.