Double Spatial Attention for Convolution Neural Networks

In recent years, attention technology has attracted a lot attention in improving the performance of convolution neural networks (CNNs). However, most existing methods mainly consider the global information including spatial and channel dimensions, which ignore the selectively augmenting the key features like human eyes. This paper proposes an efficient Double Spatial Attention (DSA) mechanism, which does not involve convolution operation, avoiding non-linear operation. DSA includes two parts: (1) calculating the spatial attention coefficients of foreground and background respectively; (2) dynamic pruning the redundant spatial pixels according to foreground and background attention coefficients. The proposed DSA module not only can effectively focus on the main features in feature maps to improve the performance, but also maximally remove the spatial feature map redundancy and reduces the redundant computation. Experiments show that our method could improve 0.03%~2.49% accuracy on MobileNetV2 without increasing additional parameters.