MLAFE-Net - Multi-Layer Attention and Feature Extraction for Change Detection
S Vishwesh Adiga, Srikrshna Parthasarthy, Rapaka Vivek, Apoorva Sarvade, S Natarajan · 2024
Change Detection (CD) is an essential task in remote sensing. With the rising popularity of Deep Learning models and image processing techniques, we intend to study and detect the changes and produce a change map. The current change detection methods fail to accurately identify the entire change region and struggle to precisely pinpoint the boundaries of the changed areas. To tackle the identified issues, we propose MLAFE-Net. In our approach we first fuse the features of the bi-temporal satellite images, resulting in the extraction of features, even low-level features. Then, to these resulting low-level features we assign weights, based on their importance. In addition to these two modules is the attention module, which captures both spatial and channel features enhancing the details captured. Direct combination of change maps from each layer facilitates effective communication, avoiding interference from indirect information. The aggregated feature maps undergo further refinement through pooling and convolution operations, employing a pyramid structure for feature fusion. This gradual integration ensures the preservation of original semantic information and incorporates diverse semantic features from various layers, yielding detailed and contextually rich information for accurate change detection. To validate and test our model we have trained and tested it on CDD and LEVIR-CD Dataset.