Blind image quality assessment (BIQA) of User Generated Content (UGC) suffers from the range effect, which indicates that on the overall quality range, mean opinion score (MOS) and predicted MOS (pMOS) are well correlated, but when focusing on a particular narrow range, the correlation is lower. To tackle this problem, a novel method is proposed from coarse-grained metric to fine-grained prediction. Concretely, we utilize global context features and local detailed features for the multi-scale distortion perception. Then, to further boost the ability of fine-grained assessment, we introduce the feedback mechanism, which is in accord with Human Vision System (HVS), to perceive detailed distortions gradually. Also, two coarse-to-fine loss functions are proposed to facilitate the feedback perception progress: a rank-and-gradient loss for coarse-grained metric keeps the assessment rank and gradient consistency between pMOS and MOS; a multi-level tolerance loss following the curriculum learning strategy is proposed to make a fine-grained prediction. Both coarse-grained and fine-grained experiments demonstrate that the proposed method outperforms the state-of-the-art ones, which validates that our method effectively alleviates the range effect. The codes are available at https://github.com/huofushuo/REQA.