Harnessing edge-enhanced attention mechanisms for supernova detection in deep learning frameworks

Recent studies have shown the advantages of convolutional neural networks in the classification and detection of supernovae. In our prior work, we employed one-stage object detection frameworks to address the challenges of presupposed location and varying image sizes in supernova detection. Notably, the backbone of the object detectors naturally emphasized the edges of candidate regions in the visualized heatmap, reflecting the strategies adopted by human observers. Capitalizing on this similarity, we introduce an innovative edge attention module, tailored to prioritize the edges of candidate regions, and improved the performance of supernova detectors. In parallel, we have developed a three-channel supernova detection dataset by integrating science (current), template (reference), and difference images into a three-channel configuration. The candidates in the new dataset are more conspicuous. To assess the efficacy of our edge attention module, we conducted a series of experiments on the proposed dataset. The experimental results establish the superiority of the proposed method in detecting supernovae. Additionally, visualizations of the feature maps shows the proposed edge attention is able to reallocate weights around the candidate edges, corroborating its effectiveness.