A de-blocking algorithm and a blockiness metric for highly compressed images

Blockiness is a typical artifact in reconstructed images that have been coded by a block-based discrete cosine transform (BDCT). In highly compressed images and video, the blocking artifacts are easily noticeable as the discontinuities between relatively homogeneous regions. Many current noniterative de-blocking algorithms attempt to remove the blocking artifacts by smoothing a few pixels around the block boundaries; however, the results are not satisfactory, especially at very low bit rates. We propose a de-blocking algorithm based on the number of connected blocks in a relatively homogeneous region, the magnitude of abrupt changes between neighboring blocks, and the quantization step size of DCT coefficients. Due to its adaptability, the proposed algorithm can smooth out the blocking artifacts while keeping the strong edges and texture areas untouched. Since this algorithm is noniterative and only identifies those block pairs that actually need de-blocking, its computation cost is low. In addition, we have developed a new metric to measure the blocking artifacts in images. Through analyzing the 2N-point (N is the block size) one-dimensional DCT coefficients of the two neighboring blocks with blocking artifacts, we show that all of the even DCT coefficients of the combined 2N points are zeros (except frequency k=0). The odd DCT coefficients are proportional to the pixel value difference between these two blocks with their magnitudes almost inversely proportional to frequency k. We selected the first DCT coefficient (frequency k=1) as an indicator for the strength of blocking artifacts in the reconstructed images. For the postprocessed images, we used a weighted summation of the squared first DCT coefficient to measure their blocking artifacts. Experimental results demonstrate that our proposed de-blocking algorithm produces better results than other methods, both visually and quantitatively, while the proposed blocking artifact metric is more consistent with subjective evaluation than the peak signal-to-noise ratio.