Point source detection with fully convolutional networks

Point sources (PS) are one of the main contaminants to recovery cosmic microwave background (CMB) signal at small scales, and their detection is important for next generation CMB experiments. We develop a method (PoSeIDoN) based on fully convolutional networks detect PS in realistic simulations, we compare its performance against most used method, Mexican hat wavelet 2 (MHW2). produce simulations taking into account contaminating signals as CMB, infrared background, Galactic thermal emission, Sunyaev-Zel'dovich effect, instrumental shot noises. first set training 217 GHz train network. Then apply both PoSeIDoN MHW2 recover validating all 143, 217, 353 GHz, comparing results by estimating reliability, completeness, flux density accuracy computing receiver operating characteristic curves. In extra-galactic region with 30{\deg} galactic cut, network successfully recovers 90% completeness corresponding 253, 126, 250 mJy respectively. The 3$\sigma$ limit up 181, 102, 153 completeness. cases produces much lower number spurious respect MHW2. techniques worsen when reducing cut 10{\deg}. Our suggest that using neural very promising approach detecting PS, providing overall better dealing usual filtering approaches. Moreover, gives competitive even nearby frequencies where was not trained.