LSAM: Local Spatial Attention Module

Miao-Miao Lv, Si-Bao Chen, Bin Luo

Abstract: We focus on spatial attention weighting to improve feature representation power of convolutional neural networks (CNNs) and propose a concise and efficient spatial attention unit based on local similarity, which is termed Local Spatial Attention Module (LSAM). Spatial neighbor points likely share similar attention. A hyper-parameter is adopted to select appropriate size of spatial neighborhood for local similarity. LSAM can be easily embedded in existing CNNs for joint end-to-end training and the cost of consumed resources is negligible. Extensive classification experiments with various backbone networks demonstrate the effectiveness of the proposed LSAM.