Score-based Aggregation for Attention Modules in Image Classification Tasks

Abstract

Deep Convolutional Neural Networks (CNNs) have been widely used for various computer vision tasks because they hierarchically extract bountiful features from a highdimensional image. Also, some CNNs incorporate channel attention mechanisms that re-scale each channel of intermediate feature maps based on their importance. The channel attention modules squeeze the spatial information of a feature into a representative value to transform it as a re-scaling value. In order to reduce the amount of information, attention modules have utilized hand-designed pooling functions such as max pooling or average pooling which have been widely adopted in CNNs, because they add negligible computational complexity. However, a significant amount of spatial information is lost due to these pooling functions. In this paper, we propose a generalized pooling function that scales down spatial information with respect to the importance of each pixel. Unlike max pooling or average pooling, our score-based aggregation is capable of flexibly adjusting to input. Also, the score-based aggregation function learns how to squeeze the spatial information into the must appropriate representative value, which will convert the pooling into a spatial attention mechanism. Finally, we propose a novel method called Score-based Aggregated Attention Module (SAAM) that utilizes the proposed score-based aggregation. Our experimental results on CIFAR-10 and CIFAR-100 datasets demonstrate that SAAM achieves the highest classification accuracy improvement among existing channel attention modules since the score-based aggregation in SAAM is a more dynamic and effective method than the hand-designed aggregations.