| Name: | Srinadh Reddy Bhavanam |
| Affiliation: | Indian Institute of Technology, Hyderabad |
| Conference ID : | ASI2023_130 |
| Title : | Combined Dictionary Learning and Attention Augmented Deep Models for Cosmic Ray Detection in Astronomical Images |
| Authors : | Dr. Sumohana S. Channappayya, Dr. P.K. Srijith, Dr. Shantanu Desai |
| Mode of Presentation: | Poster |
| Abstract Category : | Instrumentation and Techniques |
| Abstract : | Cosmic Ray (CR) hits are the primary contaminants in astronomical images obtained through optical photometric surveys. Detection and flagging CR-induced pixels from single exposure Charged Coupled Device (CCD) images is essential to ensure only top-quality data is used for scientific analysis. Both conventional and recent deep-learning-based CR detectors require experimental parameter tuning for different instruments or large volumes of training data, respectively. In this work, we present a Signal Processing algorithm called Dictionary Learning (DL) for CR detection that can embed the spatial signature of the CR hits (typically appear in patterns like dots, lines and curves) mapped in the images. Unlike the pixel-level classification, we proposed patch-level classification to detect the CR hits. We then studied how patch-level classification helps guide the pixel-level CR detection models. We characterise the CR patches uniquely using their sparse representations obtained from a learned dictionary to distinguish them from other actual astrophysical sources in the image due to the distinct and identifiable spatial signatures of CR hits. We demonstrated the performance of the proposed DL-based CR detectors using images from the Dark Energy Camera (DECam, ground-based) and Hubble Space Telescope (HST, space-based). The three main contributions in this work are 1) Novel DL-based algorithm is proposed for CR detection at the patch level, which helps obtain the coarse CR maps. 2) Coarse map generated from the DL algorithm is fed through a separate channel along with the contaminated image to detect the CR-induced pixels more accurately and better guide the deep-learning-based CR segmentation models. 3) Channel-Attention between the contaminated images and the coarse CR map is proposed to provide additional supervision to the deep models. We demonstrate that compared to the baseline deep-learning-based CR detectors, the proposed DL-augmented and DL+Attention augmented deep-learning models perform superior.
|
