Low Bit Rate Compression of Facial Images Based on Adaptive Over-Complete Sparse Representation

Among transform-based image compression methods, the sparsity of transform coefficients is very important for compression performance. To overcome the insufficiency of commonly used DCT and Wavelet transform, we apply the theory of adaptive over-complete sparse representation to the filed of facial image compression. By using a novel dictionary design algorithm called K-LMS, which recently proposed by our group, we obtain the adaptive over-complete dictionary firstly. The facial image then can be achieved sparse decomposition by using the OMP algorithm over the obtained adaptive dictionary. Finally, we encode the sparse coefficients by use of the Huffman coding. The experimental results demonstrate that the proposed method is much better than JPEG and JPEG2000 in both objective performance and visual quality, especially in the low bit-rate case.