A Dislocated 2*2 Convolution for Extracting Longitudinal and Transverse Features
Shijie Chen, Xin Xiang, Jun Zhang, Wei He · 2024
Odd-sized convolution kernels are popular in Convolutional Neural Network(CNN) today. But the even-sized kernels are rarely used. The main issue is that even-sized kernels make padding difficult when we want to keep the feature map size the same to increase the network depth. To solve this problem, we propose a dislocated 2*2 convolution, this approach solves the padding issue without adding more parameters or computations. In addition, we achieve the receptive field of a 3*3 convolution using a 2*2 convolution, this significantly improves the potential of even-sized kernels. Our convolution module can be added to any convolutional structure to help improve the feature extraction ability of the model. And from our theoretical analysis and experimental results, our module achieved the best experimental results, and it has a significant effect on extracting longitudinal and transverse features.