This paper proposes a new image compression approach that uses adaptive DCT-domain downsampling to reduce high frequency information in images for compression, and then uses learning-based mapping to compensate for the removed high frequencies during decompression. Specifically, it adaptively selects regions for downsampling in the DCT domain based on rate-distortion optimization. It then uses a database of visual patterns to map blurred image patches to corresponding high-quality patches during decompression, recovering lost high frequencies. Experimental results show it outperforms standards like H.264 and JPEG2000 especially at low bit rates.