Most image interpolations only consider the quality degradation of low-resolution images by down sampling without accounting for quantization noise. We propose a novel compression image interpolation Method using adaptive symmetrical autoregressive models in this paper. Blocks with similar local images are assumed to have the same interpolation model.The proposed Method has two phases: training and reconstruction. In the training phase

the local gradient direction is first obtained via principal component analysis (PCA) to classify all training blocks into four directions and to build the corresponding symmetrical autoregressive models and training sets for every direction. Second

the training sets for every direction are classified into subclasses according to the basic features of the K-means clustering algorithm. Finally

the model of the direction to which each subclass belongs is chosen

and the constrained least square Method is used to estimate the weights of the model. In the reconstruction phase

the direction of the pixel is first determined according to the local gradient direction of the neighboring pixel. Subsequently

by computing for the Euclidean distance between local primitive features and every clustering center in the selected direction

the model of each subclass with the least distance is chosen for interpolation. The eight test images and two quantization parameters of JPEG still image compression are used in the tests. Results show that the proposed Method is better than other interpolations on the PSNR and SSIM

even in the presence of serious quantization noise. The Results demonstrate that the proposed Method produces better Results than other interpolations for both quantitative and visual comparisons and has low computational complexity.