Facial adaptive BM3D filter: A method for compressed face image deblocking

To deal with the blocking artifacts in compressed face images, a facial adaptive BM3D (Block Matching and 3D filtering) is proposed in this paper. BM3D is one of the typical image denoising methods, which is also efficient to image deblocking. However, the specific parameters are closely related to its performance. On the observation that faces are featured with the symmetric and similar geometry structures, the BM3D can be further improved for face images deblocking. We extend the searching region of patches according to the facial symmetry. The parameter sigma of the filter is also adaptively adjusted according to the location of eyes, nose, mouth and the compressed quality. Experiments demonstrate that the proposed method can achieve better results in both the subjective and objective performance.