Identification of astronomical temporal sources in ultra-wide angle high frame images

Transient sources are important research objects in the field of time-domain astronomy, where nearby celestial bodies or moving objects such as asteroid fragments are suspected transient sources in observations. In the current stage of observation and analysis, the primary method for studying optical transient sources involves capturing multiple consecutive frames of the same sky using a computer and creating a template image using specific algorithms. This template image reflects the state of the sky over a period of time and can be subtracted from any detected image captured at a given moment to produce a residual image. Subsequently, all suspicious target images in the residual image are automatically extracted and presented to the personnel for manual identification, distinguishing between transient source phenomena and false positives. However, this method introduces human factors, leading to low operational efficiency, slow identification speed, and ultimately, the accuracy of identification is affected by the experience of the observer.To address these issues, with the rise of neural networks in recent years, convolutional neural networks (CNNs) can be trained to replace manual identification of transient sources in astronomical images. By feeding continuous astronomical images captured over time into the CNN, the network automatically identifies whether transient sources are captured in the images. Additionally, the CNN can eliminate false positives caused by errors in methods such as image alignment and subtraction. Ultimately, achieving accuracy comparable to human recognition. This project undertakes the following tasks:1. Preparation of training datasets, including two categories: positive samples of actual transient sources and negative samples of false positives caused by errors in data processing. Specific classifiers are trained to distinguish these false positives.2. The main task of this project is to build a classifier based on convolutional neural networks. After establishing the neural network structure based on previous research results, the network structure is further optimized and improved based on the characteristics of the project data, identification requirements, application scenarios, and other factors. The ultimate achievement of the project is to train an excellent transient source classifier, which can achieve an identification accuracy of over 99% when applied in actual observations.