Vision-based　outdoor　systems　are　highly　susceptible　to　severe　weather　such　as　haze.　The　quality　of　the　collected　images　in　the　hazy　environments　is　seriously　degraded,　which　affects　subjective　perception　and　brings　challenges　to　the　subsequent　intelligent　processing　tasks.　In　recent　years,　deep　learning　has　been　applied　to　single　image　dehazing　and　achieved　promising　results.　However,　the　hazy　scenes　are　complex　and　unpredictable,　which　puts　a　high　demand　on　the　generalization　ability　of　the　dehazing　methods.　In　this　paper,　we　summarize　the　recent　deep-learning-based　single-image　dehazing　methods.　The　advantages　and　disadvantages　of　these　methods　are　analyzed　in　terms　of　network　mapping　relationships,　learning　methods,　training　datasets,　and　knowledge　transfer.　In　particular,　we　focus　on　new　training　strategies　and　network　structures　that　have　emerged　in　the　last　few　years,　such　as　meta-learning,　few-shot　learning,　domain　adaption,　and　Transformer.　In　addition,　the　subjective　and　objective　performances　of　various　representative　dehazing　methods　are　compared　on　several　public　datasets.　Further,　the　impact　of　the　dehazed　images　on　the　performance　of　subsequent　object　detection　tasks　is　analyzed　and　evaluated　comprehensively.　We　also　provide　the　computational　complexity　and　running　time　of　these　methods.　Finally,　the　conclusions　and　future　tendency　of　single-image　dehazing　are　drawn.　©　2023　Chinese　Institute　of　Electronics.　All　rights　reserved.