Centrifugal and Aggregation Module Based on Shallow Features

Conventional image classification and detection tasks often rely on stacking multiple convolutional layers to increase the receptive field and obtain better features for the next step. Even with the addition of attention mechanisms, the more convolutional layers there are, the less positional information is contained in the resulting feature maps, leaving only semantic information in the end. Since the application of transformers in computer vision (e.g., Vision Transformer), many researchers have started to pay attention to the importance of positional information. Thus, we propose a centrifugal aggregation module, which aims to extract useful information from shallow features, transforming the image to the frequency domain through 2D Fourier transform, calculating its spectrum and phase maps, and then processing them through various methods before transforming back to the spatial domain. The experimental results show that adding our module to the shallow layers of most convolutional neural networks can improve the model’s performance.