As　a　paradigm　of　spontaneous　Brain-Computer　Interface　(BCI),　motor　imagery　electroencephalogram　(EEG)　has　always　been　a　hot　topic　in　BCI　and　clinical　rehabilitation.　Many　algorithms　have　been　proposed　for　decoding　the　motor　imagery　signals.　The　algorithm　based　on　the　convolutional　neural　network　has　shown　excellent　potential　in　the　task　of　motor　imagery　signal　classification.　However,　the　existing　models　are　not　specific　to　the　characteristics　of　motor　imagery　signals,　and　so,　they　cannot　fully　extract　the　signal　features　of　different　rhythms.　Moreover,　limited　by　difficulties　in　the　acquisition,　the　classification　effect　of　the　model　is　network　based　on　mixed-size　convolution　kernel　is　designed.　The　time-frequency　graph　obtained　by　Short-time　Fourier　transform　(STFT)　is　used　as　input,　and　Deep　Convolutional　Generative　Adversarial　Networks　(DCGANs)　is　used　for　data　enhancement.　The　results　show　that　the　average　classification　accuracy　is　85.7%.　Compared　with　current　mainstream　classification　algorithms,　the　model　presented　in　this　paper　has　shown　high　classification　accuracy　and　good　robustness.　©　2021　ACM.