Current　CNN-based　image　quality　assessment　(IQA)　metrics　treat　all　image　patches　equally.　We　observed　that　when　the　distortions　do　not　appear　or　not　noticeable　in　the　image　patch,　its　predicted　quality　score　is　usually　away　from　the　subjective　score.　In　this　paper,　we　argue　that　not　all　image　patches　equally　matter　for　the　IQA　task　and　present　a　multi-task　convolutional　neural　network　(CNN)　model　with　uncertainty　and　probability　for　no-reference　IQA　(MTUP).　We　also　present　a　probability-aware　average　(PAA)　method　for　pooling　the　image　quality-score　from　the　scores　of　the　image　patches.　Because　the　loss　functions　of　image　quality-score　prediction　and　distortion-type　identification　tasks　have　quantities　with　different　units　and　scales,　we　further　use　a　multi-task　learning　method　that　automatically　weights　the　loss　functions　of　multiple　tasks　with　task　uncertainty　to　solve　the　challenge　of　relative　weights　tune.　Experimental　results　show　that　the　proposed　MTUP　model　achieved　a　superior　performance　as　compared　with　state-of-the-art　NR-IQA　metrics.　©　2021　ACM.