P.1 Yang Xu (National Astronomical Observatories, Chinese Academy of Sciences)
A multi-size convolutional neural network for the recognition of optical transient
In the era of astronomical big data, the time domain survey projects can generate huge amounts of transients every day, and accurately finding astronomical flares from those transients will determine the success of those projects. In this paper, a multi-size convolutional neural network is applied to the recognition of transient; this network simultaneously extracts the features from both large and small size images of the target, which effectively solves the problem of the high false recognition rate caused by the large size difference of different categories of targets in the existing algorithms. By comparing the performance of multi-size and single-size convolutional neural networks, the test results show that the recognition performance of multi-size convolutional neural network is better than single-size convolutional neural network. The recognition accuracy is as high as 100% when use multi-size convolutional neural network to identify 1000 large-size samples labeled by hands. The trained model was applied to the actual GWAC processing flow, and the recognition rate of large-size stellar residual target was nearly 100%. Within half a month after the launch, 17 targets were found，which provides strong support for the realization of scientific objectives.
