1. EMA introduction
Paper: https://arxiv.org/abs/2305.13563v1
Acceptance: ICASSP2023
Modeling cross-channel relationships through channel dimensionality reduction may have side effects on extracting deep visual representations. This paper proposes a new efficient multi-scale attention (EMA) module. With the goal of preserving information on each channel and reducing computational overhead, some channels are reshaped into batch dimensions, and channel dimensions are grouped into multiple sub-features, so that spatial semantic features are evenly distributed in each feature group.
A new efficient multi-scale attention (EMA) without dimensionality reduction is proposed. Note that here only two convolution kernels will be placed in parallel sub-networks respectively. One of the parallel subnetworks is a 1x1 convolutional kernel, processed in the same way as CA, and the other is a 3x3 convolutional kernel. To demonstrate the generality of the proposed EMA, detailed experiments are given in Section 4, including results on CIFAR-100, ImageNet-1k, COCO and VisDrone2019 benchmarks. Figure 1 presents the experimental results on image classification and object detection tasks. Our main contributions are as follows:<