An efficient mixed attention module
Kuang Sheng, Pinghua Chen
Abstract
Abstract Recently, the application of attention mechanisms in convolutional neural networks (CNNs) has become a hot area in computer vision. Most existing methods focus on channel attention or spatial attention. Some mixed attention usually achieves better performance than channel attention or spatial attention with the help of a complex model structure, which increases the complexity of the model. This article proposes an efficient mixed attention that combines channel information with spatial information using learnable broadcast addition to reduce this complexity. In particular, this module can simplify learning and improve performance with fewer parameters. Furthermore, our method uses an excitation method based on the Tanh function to reduce computational resources while maintaining model performance, and it is a lightweight attention module that can be used in arbitrary CNNs to improve performance. Experiments on ImageNet and Cifar confirm the effectiveness of the proposed method. Besides, our method remains highly competitive for object detection tasks and image segmentation tasks.