To further improve the compression efficiency of HEVC intra frame coding, in this paper, a deep learning-based framework is proposed. Inspired by recently developed deep learning models for image super-resolution (SR), we propose to train a CNN (convolutional neural network) model to precisely predict the residual information of each CTU (coding tree unit) at the HEVC encoder. As a result, better CTU reconstruction and better prediction for the compression of subsequent CTUs can be achieved. To reduce computational complexity, different from current CNN-based SR works, we propose to skip the non-linear mapping layer, and incorporate the residual learning to obtain better predicted residual for CTU encoding. Experimental results have shown that the proposed method achieves 3.2% bitrate reduction in average BDBR (Bjentegaard delta bit rate) with only 37% encoding complexity increased.