Star-Image Centering with Deep Learning II: HST/WFPC2 Full Field of View

Dana I. Casetti-Dinescu,Roberto Baena-Galle,Terrence M. Girard,Alejandro Cervantes-Rovira,Sebastian Todeasa

2024-04-26

Abstract:We present an expanded and improved deep-learning (DL) methodology for determining centers of star images on HST/WFPC2 exposures. Previously, we demonstrated that our DL model can eliminate the pixel-phase bias otherwise present in these undersampled images; however that analysis was limited to the central portion of each detector.

Instrumentation and Methods for Astrophysics

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to solve the problem of inaccurate star center positioning in the images taken by the Wide Field Planetary Camera 2 (WFPC2) on the Hubble Space Telescope (HST). Specifically, the paper focuses on eliminating the pixel - phase bias, which can reach 40 micro - pixels (mpix) when using traditional center - positioning algorithms. This bias is mainly caused by the mismatch between the actual Point Spread Function (PSF) and the fitted PSF used in the center - positioning algorithm.

To solve this problem, the author proposes and improves a deep - learning (DL) - based method to improve the accuracy of star center positioning in WFPC2 images. Specific improvements include:

1. **Global Position and Magnitude Effects**: Global positions are introduced to compensate for PSF changes across the entire chip, and magnitude is introduced to consider non - linear effects such as charge transfer efficiency.

2. **Dataset Expansion**: A unique observational dataset of the 47 Tuc globular cluster is used, which includes offset exposure images of more than 600 different filters (F555W and F814W) with small pixel - level offsets, helping to better characterize the pixel - phase bias.

3. **Model Structure Optimization**: A new Convolutional Neural Network (CNN) model is developed, which is based on the VGG architecture and finds the best combination by automatically searching the hyperparameter space, thereby improving the model performance.

Through these improvements, the new DL model can provide more accurate star center positioning across the entire field of view of WFPC2, with an error of approximately 8 - 10 micro - pixels (mpix), and completely eliminates the pixel - phase bias.

### Formula Representation

The formulas and values involved in the paper are represented in Markdown format as follows:

- Amplitude of pixel - phase bias: $\sim 40 \, \text{mpix}$
- Correction amount corresponding to PSF change: $\sim 100 \, \text{mpix}$
- Magnitude effect level: $\sim 10 \, \text{mpix}$

These improvements enable WFPC2 images to be used for high - precision astrometry, especially for proper motion studies, thus exploiting a large amount of under - utilized archival image data.