In this paper, we propose a practical image watermark method using multi-level decoding keys. The advantages of this method are that the multiple original images are decrypted by using multi-level decoding keys in the same watermark image and that the quality of reconstructed images are clearly enhanced based on the idea of Walsh code without any side lobe in the decoding process. The zero-padded original images, multiplied with random-phase patterns to each other, are Fourier transformed and their real-valued data denote encoded images in the embedding process. The multiplication between the spreaded encoded image and each of Walsh code image is used as a hidden image. Here, the spreaded hidden image is the same size and shape of the walsh code. A stego image is then made from the linear superposition of the weighted hidden image and a cover image. Each of multi-level decoding keys is obtained by multiplying an imaginary part of the encoded image with Walsh code. The original image is simply reconstructed by the despread process of the product of the stego image and decoding key and its inverse-Fourier transform. The embedding process and the reconstruction process are performed digitally. Computer simulations are demonstrated that the efficiency of the proposed technique with multi-level decoding keys and a good robustness to the external attacks such as cropping.