| Abstrakt: |
Change detection (CD) uses images with different temporal resolutions to identify changes occurred over a period of time. Recently, various deep learning based frameworks are used for CD in remote sensing community due to its stellar performance. Most of the active deep learning pipelines are dependant on two mechanism namely, CAM (channel attention mechanism) and SAM (spatial attention mechanism) for capturing long-range dependencies to derive different feature characterization. However, both CAM and SAM are employed independently in sequential or parallel manner. Therefore, they do not serve any relationship and do not achieve cross dimension linking of channel dimension (c) with spatial dimensions (hand w). Moreover, any imbalance between changed feature and unchanged feature might result in pseudo-changes. As a consequence of these reasons, CD methods based on attention module do not obtain discriminate feature representation. To get over this problem, we formulated a building CD algorithm which is established on Siamese convolutional neural network (Deep-SiamCNN) for feature extraction followed by two-stream modified triplet attention mechanism (TSMTAM) to enhance the feature extracted from Deep-SiamCNN network which construct cross-dimension interactions between c, h, and w. The proposed TSMTAM composed of three parallel structures out of which two parallel structures are utilized for captivating cross dimension dependencies between cand hor wwhile the third parallel structure is accountable for constructing spatial dependencies. To resolve the hurdle of imbalanced feature pairs, we framed-up dual-margin contrastive loss function to isolate changed feature pairs while drag together unchanged feature pairs. Two standard datasets namely LEVIR-CD and WHU building were used for experimentation. The experimental results on both dataset demonstrated virtue of the proposed building CD pipeline and exhibited promising results over traditional building CD algorithms. |
Full text from SpringerLink