Haze　and　fog　weather　decrease　the　accuracy　of　license　plate　recognition　using　machine　vision.　To　address　this　issue,　the　paper　proposes　a　multi-stage　deep　learning-based　license　plate　recognition　algorithm　with　fog　detection　and　defogging.　The　algorithm　utilizes　an　image　average　gradient-based　fog　detection　model　and　an　adaptive　defogging　processing　approach　with　the　improved　ACE　algorithm＇s　adaptive　color　balancing　fast　algorithm　to　enhance　the　license　plate　region＇s　visibility.　Furthermore,　image　recognition　and　segmentation　theory,　edge　detection　algorithm,　HSV　color　model,　and　mathematical　morphology　algorithm　are　applied　to　preprocess　the　license　plate　area.　The　convolutional　neural　network　is　then　utilized　to　filter　the　license　plate,　enabling　efficient　license　plate　location　and　extraction.　Finally,　another　convolutional　neural　network　is　employed　for　character　recognition.　The　result　shows　a　recognition　accuracy　of　98.7%　in　hazy　weather　and　99.3%　in　non-fog　conditions　with　fast　computation　speed.