Abstract: A block-adaptive image compression algorithm based on dictionary learning is proposed for images with clear structures such as cartons. The KSVD algorithm was used to train an offline dictionary for the carton image sample set, the OMP algorithm was used to calculate the initial sparse coefficient matrix of the image block under the original error, the improved Canny edge detection was used to determine the structure complexity of the image block and adaptively set the sparse representation model's quadratic optimization error for block partitioning, to achieve targeted compression of different image blocks. The experimental results of multiple carton images with different text density show that this algorithm can achieve a compression ratio of 0.95% to 1.95% while preserving the important information of the image.