关键词: 特征融合, 残差网络, 卷积神经网络, 大核卷积

Abstract: As the most widely adopted backbone network in classification, object detection and instance segmentation tasks, the representation capability of ResNet ( Residual Neural Network) has gained extensive recognition. However, there are still certain limitations that hinder the representation ability of ResNet, including feature redundancy and inadequate effective receptive field. To address these problems, a feature fusion block is proposed, which can fuse features of different scales to construct multi-scale features with richer information and improve channel utilization, when the model channel is increased. The block employs a small number of large kernel convolutions, which is benefit to the expansion of the effective receptive field of the model and the trade-off between performance and computational efficiency. And a lightweight downsampling block and a shuffle compression block are also proposed, which can effectively reduce the parameters of the model and make the entire method more efficient. The feature fusion block, downsampling block and shuffling compression block are introduced to the ResNet can build a FFNet(Feature Fusion Network), which will have faster convergence speed and a larger effective receptive field and better performance. Extensive experimental results on CIFAR (Canadian
Institute for Advanced Research ), ImageNet and COCO ( Microsoft Common Objects in Context ) datasets demonstrate that the feature fusion network can bring significant performance improvements in classification, object detection and instance segmentation tasks while only adding a small number of parameters and FLOPs(Floating Point Operations).

Key words: feature fusion, ResNet, convolutional neural network, large kernel convolution