Abstract

Nowadays, the tidal waves of deep convolution have promoted the proliferation of deep learning change detection (CD) methods. However, challenges still remain as most algorithms tend to have poor detections of small targets, unsmooth edges, and incomplete internal regions, largely because of a lack of effective features, context information, and feature fusion. In this paper, a multi-attention feature-constrained pixel-shuffle image fusion network (MapsNet) is proposed to address the challenges in CD tasks. We first employ a two-stream fully convolutional network for feature extraction, which is adaptively constrained by the proposed multi-attention module (MAM). The capability of the MAM module is further enhanced by the introduction of a novel attention module, i.e., CBAM-s. In addition, we propose a pixel-shuffle image fusion network (PSIFN) to aggregate multi-level contextual information and implement feature fusion to complete change map reconstruction. The conducted experimental results confirm that the MapsNet demonstrates better effectiveness and robustness in complex changing scenes compared to selected state-of-the-art CD methods. Finally, these novel strategies designed in this article have high practicability and transferability.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call