Improving the Performance of Convolutional Neural Networks by Fusing Low-Level Features With Different Scales in the Preceding Stage

Abstract

The width of convolutional neural networks (CNNs) is crucial for improving performance. Many wide CNNs use a convolutional layer to fuse multiscale features or fuse the preceding features to subsequent features. However, these CNNs rarely use blocks, which consist of a series of successive convolutional layers, to fuse multiscale features. In this paper, we propose an approach for improving performance by fusing the low-level features extracted from different blocks. We utilize five different convolutions, including 3×3, 5×5, 7×7,5×3 ∪ 3×5 and 7×3 ∪ 3×7, to generate five low-level features, and we design two fusion strategies: low-level feature fusion (L-Fusion) and high-level feature fusion (H-Fusion). Experimental results show that the L-Fusion is more helpful for improving the performance of CNNs, and the 5×5 convolution is more suitable for multiscale feature fusion. We summarize the conclusion as a strategy that fuses multiscale features in the preceding stage of CNNs. Furthermore, we propose a new architecture to perceive the input of CNNs by using two self-governed blocks based on the strategy. Finally, we modify five off-the-shelf networks, DenseNet-BC (depth = 40), ALL-CNN-C (depth = 9), Darknet 19 (depth = 19), Resnet 18 (depth = 18) and Resnet 50 (depth = 50), by utilizing the proposed architecture to verify the conclusion, and these updated networks provide more competitive results.