Multiresolution analysis based sparse dictionary learning for remotely sensed image retrieval

Sparse representation using over-complete dictionaries offers a compact representation with minimal error and yields in compression ratios that are relatively higher than other conventional lossless compression algorithms. Multiresolution analysis (MRA) offers a framework to decompose and observe a signal at various resolutions thus allows representing the signal in a way that makes the retrieval convenient.In this paper, we propose a wavelet based MRA framework for image compression by sparse representation of the decomposition of an image, trained on over-complete dictionary. The dictionaries are trained using algorithms like the singular value decomposition (SVD), method of optimal directions (MOD) and the sparse representation is done using the pursuit algorithm. The method is tested on remotely sensed images captured by Worldview-1 satellite and reconstructions from the dictionaries using proposed framework also yield a low root mean square error and a high spatial similarity index (SSIM) which attests to the high quality of reconstructed images. The results obtained show the storage of the over-complete dictionaries trained on the various subbands, keep on reducing as the level of decomposition increases lead to a better compression ratio than the sparse representation of an image at a single resolution while maintaining the same image quality.