基于联合稀疏表达图像融合.pdfVIP

IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 0, NO. 0, 2011 1 Image Features Extraction and Fusion Based on Joint Sparse Representation Nannan Yu, Tianshuang Qiu, Feng Bi, and Aiqi Wang Abstract— In this paper, a novel joint sparse representa- of the source image in some bases [such as the discrete Fourier tion-based image fusion method is proposed. Since the sensors transform (DFT), the mother wavelet and cosine bases of the observe related phenomena, the source images are expected to pos- discrete cosine transform (DCT)] are used as image features in sess common and innovation features. We use sparse coefficients as image features. The source image is represented with the common the transform domain method. The fusion rules mainly include and innovation sparse coefficients by joint sparse representation. “choose-max” and “weighted average.” The wavelet-based The sparse coefficients are consequently weighted by the mean image fusion method “wavelet maxima” [4] selects the wavelet absolute values of the innovation coefficients. Furthermore, since coefficients with the largest activity level at each pixel location. sparse representation has been significantly successful in the Mahbubur Rahman et al. [2] calculate the weighted mean of development of image denoising algorithms, our method can carry out image denoising and fusion simultaneously, while the images wavelet coefficients of the source image for image fusion, are corrupted by additive noise. Experiment results show that the and the weights depend on the contrast of the source images. performance of the proposed method is better than that of other The authors in [5] and [6] use the mean absolute values of the methods in terms of several metrics, as well as