Dense Attention Fusion Network for Object Counting in IoT System

Abstract

IoT has been overwhelmingly empowered by the rapid development of big-data ecosystems, such as remote sensing technology which runs all the time in obtaining accurate and high-quality images to facilitate the subsequent image processing and content analysis in embedded devices. Object counting, which aims to estimate the number of objects in a captured image, is one of the most crucial tasks among multimedia data and wireless network. However, there are enormous inherent factors that seriously degrade the counting performance in remote sensing, e.g. the background clutter, scale variation, and orientation arbitrariness. In this paper, we tackle the aforementioned problems in a divide-and-conquer manner by devising the dense attention fusion network (DAFNet). Specifically, we introduce an iterative attention fusion (IAF) module, which mainly relies on the multiscale channel attention (MCA) unit, to alleviate the side effect caused by background clutter. Meanwhile, to overcome the intrinsic scale variations, we build a dense spatial pyramid (DSP) module to consider the hierarchical information obtained under diverse receptive fields. Finally, we stack deformable convolution layers to deal with the orientation arbitrariness. The synergy of the proposed IAF and DSP modules substantially promotes the effectiveness of the proposed DAFNet, which can be demonstrated by the notable superiority in extensive experiments on the remote sensing counting datasets against state-of-the-art competitors.