Pavement　crack　detection　has　been　a　challenging　yet　useful　task　in　computer　vision,　due　to　the　complexity　of　cracks　in　the　road　surface.　In　this　paper,　we　thoroughly　analyze　the　role　of　context　information　to　extract　both　local　and　global　features　for　crack　detection.　Concretely,　we　observe　that　pavement　crack　detection　requires　local　context　information　at　the　low-level　feature　extraction　stage.　Leveraging　the　dilated　convolution　to　achieve　a　large　receptive　field,　we　extract　multi-scale　context　information　by　setting　different　dilated　rates.　The　proposed　unconventional　feature　extraction　module　is　simple,　yet　effective　and　efficient　for　crack　detection(UFE-net).　By　comparing　with　those　state-of-the-art　methods,　experimental　results　demonstrate　our　method　achieves　the　best　results　with　an　AP　87.48%.　©　2022　IEEE.