Deep　neural　network　is　a　promising　method　to　recognize　motor　imagery　electroencephalography　(MI-EEG),　which　is　often　used　as　the　source　signal　of　a　rehabilitation　system;　and　the　core　issues　are　the　data　representation　and　the　matched　neural　networks.　MI-EEG　images　is　one　of　the　main　expressions,　however,　all　the　measured　data　of　a　trial　are　usually　integrated　into　one　image,　causing　information　loss,　especially　in　the　time　dimension;　and　the　neural　network　architecture　might　not　fully　extract　the　features　over　the　alpha　and　beta　frequency　bands,　which　are　closely　related　to　MI.　In　this　paper,　we　propose　a　key　band　imaging　method　(KBIM).　A　short　time　Fourier　transform　is　applied　to　each　electrode　of　the　MI-EEG　signal　to　generate　a　time-frequency　image,　and　the　parts　corresponding　to　the　alpha　and　beta　bands　are　intercepted,　fused,　and　further　arranged　into　the　EEG　electrode　map　by　the　nearest　neighbor　interpolation　method,　forming　two　key　band　image　sequences.　In　addition,　a　hybrid　deep　neural　network　named　the　parallel　multimodule　convolutional　neural　network　and　long　short-term　memory　network　(PMMCL)　is　designed　for　the　extraction　and　fusion　of　the　spatial-spectral　and　temporal　features　of　two　key　band　image　sequences　to　realize　automatic　classification　of　MI-EEG　signals.　Extensive　experiments　are　conducted　on　two　public　datasets,　and　the　accuracies　after　10-fold　cross-validation　are　97.42%　and　77.33%,　respectively.　Statistical　analysis　shows　the　superb　discrimination　ability　for　multiclass　MI-EEG　too.　The　results　demonstrate　that　KBIM　can　preserve　the　integrity　of　the　feature　information,　and　they　well　match　with　PMMCL.