This paper presents a novel coder architecture for image compression based on Independent Component Analysis (ICA). ICA has been used to extract the independent features (basis functions) from a set of images. These basis functions are localized, band-limited and oriented like human visual system (HVS) and resemble wavelet and Gabor basis. Having resemblance with curvelets, it is natural to use these basis functions for Image compression. Greedy algorithm like matching pursuit (MP) has been used to transform the image in the ICA domain which is followed by quantization and entropy coding stages. We have compared our codec with JPEG from the DCT domain and with JPEG2000 from the wavelets domain. For fingerprint images, comparison is also made with Wavelet Scalar quantization (WSQ) coder. Our proposed codec outperforms JPEG and WSQ and also performs very close to JPEG2000. Our coder presents lower complexity than JPEG2000. SNR and PSNR are used for objective quality comparison while picture quality scale (PQS) is used for subjective quality measurement as it performs very close to HVS.