| Abstract: | Filaments are key features on the solar surface. When the magnetic field destabilizes, these filaments erupt as flares and coronal mass ejections (CMEs), releasing stored plasma into space. These eruptions contribute to space weather activities, emphasizing the importance of filament detection for studying the solar magnetic field and predicting space weather events. Solar filaments appear as dark, thin, rope-like structures with low temperatures and high plasma density. They are visible in H𝛼 full-disk solar images at the 656.3 nm spectral line and in ultraviolet at the 393.3 nm Calcium K line. This in turn converts the overall task of solar filament detection into a task of identifying thin and long dark features in solar images. Over time, various supervised and unsupervised computer vision methods have been applied to detect solar filaments, such as global thresholding, local thresholding, artificial neural networks, and deep learning techniques. However, these methods are not universally effective specially for the case of solar images having non-homogeneous distribution of intensity level, on the other hand deep learning approaches, in particular, require extensive labeled datasets. Alongside, the rise in both ground and space-based solar observatories has increased the volume of solar images, necessitating efficient, real-time, and automated filament detection methods. To address this, an integral adaptive thresholding-based unsupervised approach has been developed. This method first segments solar images and extracts dark features using adaptive thresholding, followed by a disconnected component analysis to isolate filament regions from these extracted dark features. Tests on full-disk H𝛼 images from the Big Bear Solar Observatory (BBSO) in 2013 show that this approach achieves an accuracy rate above 99% for most solar images, outperforming traditional object detection algorithms. This adaptive method offers a reliable solution for automated filament detection across a range of solar images. |
