Abstract: Almost all images are compressed to be transferred or stored, exhibiting various visual artifacts. For this reason, identifying the compression quality level with only visual cues is the starting point to enhance the image quality. This paper introduces a compression quality prediction method, named Q1Net, which yields a single quality level with over 99%-accuracy in a matter of milliseconds on mobile devices regardless of the image resolution. This real-time and high-accurate performance is attributed to the observation that most image compression methods are based upon transform coding on small blocks of different characteristics. To separately investigate and exploit the distinct visual deformations induced by one transform coding, our method measures the compression quality level on various image patches containing a basic coding block and its neighboring pixels. Our approach then elaborately selects promising candidate patches that can indicate the compression quality reliably through CNN-based statistical confidence estimation. In order to make a final decision, the proposed method fuses the prediction results from a selected number of input patches, which makes it scalable and operable on mobile devices with varying computational capabilities. According to the extensive experiments on the DIV2K dataset and an off-the-shelf smartphone, our block-wise confidence-aware Q1Net achieves better performance in compression quality prediction than other well-known CNN-based methods in terms of speed and accuracy.