Based on geometric properties of the local image structures and the perception characters of human visual system (HVS), a geometrically motivated Gabor multi-component perception dictionary is constructed. Moreover, inspired from the hierarchical information processing in the human visual path, a sparse coding network is proposed to reduce high order redundancy, thus provides much sparser representations of images. Furthermore, coefficients of repositioned atoms in sparse approximation are quantized by bit-plane quantization. It presents an effective low bit-rate and scalable image compression algorithm. Our simulation results show that, under the low bit-rate, our approach is comparable to the JPEG2000 in terms of PSNR value, while effectively preserves edges and textures structures and exhibits generally a better visual quality.