Research　on　brain-computer　interfaces　(BCI)　can　identify　the　limbs　of　subjects　that　generate　motor　imagination　by　decoding　brain　physiological　signals.　The　features　extracted　from　traditional　Electroencephalography　(EEG)-based　decoding　methods　are　relatively　single　and　limited,　leading　to　the　limited　decoding　performance　of　classification　methods.　In　order　to　solve　this　problem,　we　proposed　a　complementary　feature　fusion　network　(CFFNet)　model　based　on　EEG　and　functional　near-infrared　spectroscopy　(fNIRS)　signals　since　the　complementary　information　content　between　EEG　and　fNIRS　signals.　The　CFFNet　method　integrates　the　feature　extraction　block,　the　feature　selection　block,　the　complementary　feature　fusion　block,　and　the　shared-specific　feature　for　complementary　feature　fusion　learning,　which　enables　effective　capability　of　utilizing　the　shared　and　specific　information　of　each　modal.　This　approach　and　representative　MI　recognition　methods　were　evaluated　an　open　multimodal　dataset.　The　method　we　proposed　achieved　an　average　accuracy　of　76.45%　in　intra-subject　experiments,　which　is　significantly　higher　than　single-modal　classification　methods　and　slightly　higher　than　the　representative　multi-modal　BCI　methods.　Comprehensive　experimental　results　verify　the　effectiveness　of　our　proposed　method,　which　can　provide　novel　perspectives　for　multi-modal　decoding.　©　2024　ACM.