ADCM: attention dropout convolutional module

Network architecture design plays an important role in boosting the performance of models in various applications. In this work, we design a general and lightweight module named the attention dropout convolutional module (ADCM). It consists of two submodules, channel attention dropout (CAD) and position attention dropout (PAD), and each submodule integrates both attention and dropout mechanisms. The attention mechanism emphasizes the meaningful information and suppresses unnecessary noise. The dropout-channel in the CAD submodule filters the channel based on its channel attention, while the dropout-region in the PAD submodule filters the region consisting of the spatially correlated features according to its position attention. The two dropout methods we designed allow the baseline network to learn more robust features and further boost its performance. Finally, we deploy the ADCM in consecutive layers of classical convolutional neural networks and evaluate its performance on multiple benchmark datasets. The experimental results demonstrate that the ADCM brings significant improvements to the performance of the baseline models at negligible computational cost and with less complexity. (C) 2020 Elsevier B.V. All rights reserved.