A. N. Ganin, O. N. Gushchina, V. V. Khryashchev
Block discrete cosine transform (B-DCT), computed on a square or rectangular support, has been widely used as a key element in many compression algorithms and image processing applications. In 1995 an adaptive discrete cosine transform (A-DCT) was proposed by Sikora. The A-DCT is computed by cascaded application of one-dimensional varying-length DCT transforms first on the columns and then on the rows that constitute the considered image region. The shape of the transform’s support is defined with the use of anisotropic local polynomial approximation (ALPA). The image processing algorithm is based on the sequential use of A-DCT, hard-thresholding and inverse A-DCT. Hard-thresholding is performed after noise variance approximation. The algorithm was implemented in MatLab. Results of image processing were estimated in reference to three objective image quality metrics: peak signal-to-noise ratio, structural similarity index and modified peak signal-to-noise ratio. The algorithm was applied for noise removal of Additive White Gaussian Noise and post-processing of JPEG-compressed images, and comparison with other filters (Wiener filter, bilateral filter) and deblocking algorithm (Algorithm for Blocking Removal) was demonstrated. In this paper the analysis of noise variance approximation accuracy on algorithm performance was made. The experimental results demonstrated that algorithm of image processing with the use of A-DCT outperforms other reference methods in terms of all three objective metrics. Visual examples confirmed that algorithm can be effectively used for image filtering and removal of blocking artifacts on JPEG-compressed images.