【摘要】 针对当前图像伪造检测算法难以有效识别相似区域,且其检测精度依赖于参数和阈值的选择,使其自适应能力较差等不足,提出局部二值模式耦合K-D树的复制-粘贴图像伪造检测算法。基于传统的LBP,设计具有均匀不变性、旋转不变性以及旋转均匀不变性的3种LBP,将其组合形成多尺度LBP(MLBP),利用MLBP提取图像特征,得到3组特征矩阵;引入K-D树,寻找其最优邻域,获取对应的3个相似特征矩阵,通过判断3个特征矩阵中至少两个特征值相同来识别该区域是否被篡改;引入随机抽样一致性策略,降低图像块的误匹配率,提高检测精度。实验结果表明,与当前图像伪造检测技术相比,该检测算法的检测精度更高,能有效识别出旋转、缩放、模糊以及噪声等伪造形式。更多还原

【Abstract】 For the detecting difficulties of image forgery detection algorithm in similar area,and that the detection accuracy depends on the parameters and threshold,leading to poor adaptive ability,the copy-paste image forgery blind detection algorithm based on LBP coupled K-D tree was proposed.Based on the traditional LBP,three kinds of LBP with the quality of uniform and rotational invariance and rotation invariance uniform were designed,which were combined into multi-scale LBP(MLBP).Three sets of characteristic matrixes were obtained through the image features extracted by MLBP.K-D tree method was introduced to find the optimal neighborhood to get the corresponding three similarity matrixes and to identify whether the region was forged by judging at least two of the three characteristic matrixes with the same value.RANSAC optimaization was introduced to reduce image block matching error rate and improve the accuracy of detection.Experimental results show that this algorithm can accurately detect the tampered area of rotation,zoom,blur and noise distortion with wide applicability and good robustness.更多还原