In this paper, we propose a coding method for camera raw images with high dynamic ranges. Our encoder has two layers. In the first layer, 24bit low dynamic range image is encoded by a conventional codec, and then the residual image that represents the difference between the raw image and its approximation is encoded in the second layer. The approximation is derived by a polynomial fitting. The main advantage of this approach is that applying the polynomial model reduces the correlation between the raw and 24bit images, which increases coding efficiency. Experiments shows compression efficiency is significantly improved by taking an inverse tone mapping into account.