Combining neural networks with galaxy light subtraction for discovering strong lenses in the Hyper Suprime-Cam Subaru Strategic Program

Abstract Galaxy-scale strong gravitational lenses are valuable objects for a variety of astrophysical and cosmological applications. Strong lensing galaxies are rare, so efficient search methods, such as convolutional neural networks (CNNs), are often used on large imaging datasets. In this work, we apply a new technique to improve the performance of supervised neural networks by subtracting the central (lensing) galaxy light from both the training and test datasets. We use multiband imaging data from the Hyper Suprime-Cam Subaru Strategic Program as our training and test datasets. By subtracting the lensing galaxy light, we increase the contrast of the lensed source compared to the original imaging data. We also apply the light subtraction to non-lenses in order to compare them to the light-subtracted lenses. Residual features resulting from poor light subtraction can adversely affect the performance of networks trained on the subtracted images alone. We find that combining the light-subtracted images with the original $gri$-band images for training and classification can overcome this and improve the overall classification accuracy. We find the area under the receiver operating characteristic curve can be improved to 0.841 using the combination of the fiducial images and light-subtracted images, compared to 0.808 for the fiducial imaging dataset alone. This may be a promising technique for improving future lens searches using CNNs.