Image compression through optimized linear mapping and parametrically generated features

In this paper a new linear mapping scheme for image compression is proposed. The main objective is to construct an asymptotic approximation to higher order surface fitting schemes previously reported in [1]. Each block of the image is independently reconstructed from a set of "features" through a linear mapping. These features should be (ideally) independent or at least uncorrelated to benefit the info-max principle. A random sequence generator is employed, using a sine function with two parameters, to approach such a requirement for the features. These two parameters and the set of linear weights are found through an optimization process. An off-line training phase is first performed to find the weights used in the linear mapping. These weights are then used to compress images during the on-line phase where the sine function parameters are found and quantized. The computation time is excessive due to nonlinear optimization required. However, thanks to quantization, a look-up table can be implemented to overcome this disadvantage. The proposed structure is fairly robust to the random sequence length as experimentally demonstrated.