基于稀疏表示的视网膜图像对变化检测

视网膜图像对变化检测主要研究两个不同时间点所采集到图像的变化情况。图像间的亮度差异及解剖结构与病灶的亮度相似性,使得基于逐点对比的差分方法或商方法很难准确检测变化区域。针对光照干扰问题,提出对光照具有鲁棒性的稀疏表示(SRC)变化检测方法。SRC方法先抽取参考图像局部区域块构建字典,再通过稀疏表示重构当前图像的局部背景块,最后利用背景相减获得变化区域。通过该方法,图像对的亮度差异可用稀疏表示系数自动调整,而基于块的方式可过滤掉局部光照,更有效地检测出变化区域。SRC方法与其他检测方法结合,可以增加检测结果的准确性。实验根据一对来自DRIVE数据集的小病灶仿真数据,SRC方法的AUC和mAP值分别为0.986和0.865;对一对采集自临床的大病灶数据,SRC与迭代鲁棒同态曲面拟合(IRHSF)校正结合方法的AUC和mAP值分别达到了0.989和0.969。实验结果表明,SRC方法比RPCA方法对局部光照鲁棒性更强,比基于逐点的比较差分方法更多地考虑局部邻域信息,能够更有效地检测出变化区域。

Change Detection Based on Sparse Representation for Retina Fundus Image Pair

Change detection with a pair of retinal fundus images was focused on comparing two images captured at different stages. Illumination variations between the image pair, along with the intensity similarity of anatomic structures and lesions make the pixel-by-pixel methods based on subtraction operation or ratio operation hard to obtain clear changing areas. In this paper, a new change detection method based on sparse representation classification (SRC) was proposed, aiming to reduce the illumination variations between the image pair. The SRC method first extracted the local neighborhood patches from the reference image to construct a local dictionary, then reconstructed the background of the current image by sparse representation on the extracted dictionary. Finally, change areas were obtained through background subtracting. The illumination variations between two images were corrected automatically by the representation coefficients, and SRC method based on patches can filter local contrast effectively to detect changing areas. A combination of SRC and some other change detection methods can improve the accuracy of the detection result. In the experiments of the passage, for a simulated image pair with small lesions, the AUC and mAP values were 0.985 and 0.864 respectively. For a clinical image pair with a big lesion, the AUC and mAP values of the combination of SRC and iterative robust homomorphic surface fitting (IRHSF) were 0.989 and 0.969 respectively. Experimental results showed that SRC was more robust than RPCA for the illumination variations and could detect the changing area more effectively than pixel-wised subtraction as it was involved with more neighborhood information.