In　submarine　information　acquisition,　optical　imaging　is　an　important　way　for　its　ease　of　understanding　by　humans　and　machines.　Among　all　optical　pictures,　underwater　videos　are　friendly　to　complicated　surveillance　tasks　such　as　object　tracking　and　event　analytics.　However,　the　videos　are　still　barely　used　in　real-time　oceanic　surveillance　tasks　because　the　underwater　video　transmission　is　yet　to　be　well　addressed.　The　popular　underwater　acoustic　channel　provides　a　significantly　lower　capacity　than　wired　or　wireless　channels.　How　to　support　high-resolution　video　transmission　at　such　a　limited　capacity　is　an　open　issue.　We　address　this　issue　by　proposing　a　joint　framework　of　front-end　detection　and　video　compression.　First,　we　automatically　employ　deep　neural　networks　ProposeReduce　to　automatically　identify　all　target-aware　regions.　We　also　designed　a　mask　compensation　module　to　compensate　missed　areas　for　semantic　consistency　across　object　borders.　Second,　we　deploy　a　multi-level　background　smoother　on　target-free　regions　to　guarantee　a　low-bitrate　background　compression　without　impact　on　the　integrity　of　targets.　Third,　we　feed　all　processed　frames　to　a　unified　video　encoder　for　compression.　Experimental　results　show　that　our　method　reduces　the　bitrate　of　underwater　videos　by　58.05%.