Purpose. Digital images are easy to be polluted in the communication. The research on image denoising is aimed to develop a new image denoising approach based on sparse representation, which will allow removing the noises in the digital images effectively and improve the image quality. Methodology. By using K-SVD (K-means Singular Value Decomposition) algorithm, we trained the DCT (Discrete Cosine Transform) dictionary into a new dictionary, in which every atom was a linear combination of the atoms from the original DCT dictionary. The composition of these two dictionaries differs greatly, which proves that K-SVD algorithm is able to improve the dictionary structure effectively. Findings. At first, we described and analyzed image denoising briefly and then discussed the relevant algorithms and techniques of sparse representation based on the initialization of DCT dictionary. Based on the above theories and techniques, a new image denoising method based on K-SVD and adaptive dictionary was developed. Originality. By combining the construction and optimization of over-complete dictionary, we trained the atom dictionary with the samples of the images to be decomposed so that we managed to build the atom dictionary that can effectively reflect various image features. Through simulation analysis, this noise removal method allows denoising the images with profound details and increasing the peak signal to noise ratio of the image effectively. Practical value. The image denoising method based on sparse representation has been developed. This approach makes contribution to the update steps of the dictionary and it solves the problem of matrix inversion by making iterative updates to every row of the matrix. Which is more important, this algorithm also updates the relevant coefficients while updating the atoms in the new dictionary and greatly reduces the computation complexity.