Raindrop　removal　from　a　single　image　is　an　important　low-level　computer　vision　task　for　autonomous　driving　and　object　detection.　Current　CNN-based　methods　achieve　encouraging　performance,　while　are　limited　to　single-scale　and　then　make　the　network　more　deeper　to　achieve　good　performance.　To　address　this　issue,　we　present　a　Multi-scale　Adaptive　Feature　Fusion　Network　(MSAFF-Net)　in　end-to-end　manner,　to　exactly　capture　raindrop　features　with　multi-scale　extraction　and　information　aggregation.　For　better　extracting　the　features,　a　novel　Multi-scale　Extraction　Block　(MSEB)　is　proposed　to　get　local　and　global　features　across　different　scales　through　dilated　convolution　with　different　dilation　rate.　Besides,　we　design　an　Adaptive　Feature　Fusion　module　(AFF)　to　aggregate　different　features　instead　of　directly　concatenating　or　adding.　Extensive　experiments　on　datasets　demonstrate　the　effectiveness　of　the　designed　MSAFF-Net　by　comparing　with　recent　state-of-the-art　raindrop　removal　algorithms.　©　2021　ACM.