图像哈希：Global+Local

文章信息

作者：梁小平，唐振军
期刊：ACM Trans. Multimedia Comput. Commun. Appl（三区）
题目：Robust Hashing via Global and Local Invariant Features for Image Copy Detection

目的、实验步骤及结论

目的：通过全局和局部提取特征来生成最终图像的哈希值。
实验步骤：
- 数据预处理：双线性插值（512 * 512）
- 全局特征：
  - PDFT生成显著图S
  - 对GLCM使用四种参数（不同的角度）得到四个矩阵，每个矩阵得到4个统计特征，得到 1 * 16 的全局特征向量
- 局部特征：
  - 使用HSV中的V分量，分块（64 * 64），将每一个块拼接成一个列向量，使用KPCA后得到d * N的矩阵。
  - 计算每一个矩阵维度的均值作为参考向量，计算所有向量（每一列）和参考向量的距离作为局部特征
- 生成哈希值：将全局特征和局部特征进行拼接，使用量度排序作为最后的哈希值（长度为N+16）。
- 相似性评价：使用汉明距离判断两张图片是否一致，若小于阈值则是相同图片。
结论
- 首次提出KPCA应用于图像哈希
- 适用于混合攻击
- 全局特征对几何攻击（尤其是缩放和旋转）很敏感，而局部特征无法保持全局上下文信息导致判别效果不佳。
- 使用全局和局部结合特征可以更加有利于互补进行提取特征。

本篇论文的实现代码如下：

def image_hash(img_path):img = processing(img_path)global_feature = global_feature_gen(img)local_feature = local_feature_gen(img, 10000, 4)h_i = gen_hashing(global_feature, local_feature)return h_idef processing(img_path):"""input：图片的路径output：处理后的RGB图片"""try:img = cv2.imread(img_path)x = img.shape[0]//2 # 高度y = img.shape[1]//2 # 宽度Min = x if x<y else ycropped_image = img[x-Min:x+Min, y-Min:y+Min] # 裁剪图像img = cv2.resize((cropped_image), (512,512), interpolation=cv2.INTER_LINEAR)except:img = imageio.mimread(img_path)img = np.array(img)img = img[0]img = img[:, :, 0:3]x = img.shape[0]//2 # 高度y = img.shape[1]//2 # 宽度Min = x if x<y else ycropped_image = img[x-Min:x+Min, y-Min:y+Min, :] # 裁剪图像img = cv2.resize((cropped_image), (512,512), interpolation=cv2.INTER_LINEAR)
#     out = cv2.GaussianBlur(img, (3, 3),1.3) # 使用python自带的高斯滤波kernel = np.array([[1,2,1],[2,4,2],[1,2,1]])/16out = cv2.filter2D(img, -1 , kernel=kernel)  # 二维滤波器# out = cv2.cvtColor(out, cv2.COLOR_BGR2RGB)out = cv2.cvtColor(out, cv2.COLOR_BGR2HSV)return outdef local_feature_gen(img, sigma, n_components):"""iamge:(512,512,3)return: 降维之后的图像(d, N)"""from sklearn.decomposition import PCA, KernelPCAN_list = []V = img[:,:,2]for i in range(0,V.shape[0],64):for j in range(0,V.shape[1],64):image_block = V[i:i+64, j:j+64]N_list.append(image_block.reshape(-1)[:])N_list = np.array(N_list).copy()# kernel_pca = KernelPCA(n_components=4, kernel="poly", gamma=10)# result = kernel_pca.fit_transform(N_list)result = kpca(N_list, sigma, 4).copy()return result.Tdef gaussian_kernel(X, sigma):sq_dists = pdist(X, 'sqeuclidean')  # 计算所有样本点之间的平方欧式距离mat_sq_dists = squareform(sq_dists)  # 转换成矩阵形式return np.exp(-mat_sq_dists / (2 * sigma**2))  # 计算高斯核矩阵def kpca(X, sigma, n_components):# 步骤1: 计算高斯核矩阵K = gaussian_kernel(X, sigma)# 步骤2: 中心化核矩阵N = K.shape[0]one_n = np.ones((N, N)) / NK = K - one_n.dot(K) - K.dot(one_n) + one_n.dot(K).dot(one_n)# 步骤3: 计算特征值和特征向量eigenvalues, eigenvectors = eigh(K)eigenvalues, eigenvectors = eigenvalues[::-1], eigenvectors[:, ::-1]  # 降序排列# 步骤4: 提取前n个特征向量alphas = eigenvectors[:, :n_components]lambdas = eigenvalues[:n_components]return alphas / np.sqrt(lambdas)  # 归一化特征向量def global_feature_gen(img):P = pqft(img)return P
def pqft(img, sigma=8):h, w, channel = img.shaper, b, g = img[:,:,0], img[:,:,1], img[:,:,2]R = r - (g + b)/2G = g - (r + b)/2B = b - (r + g)/2Y = (r + g)/2 - (abs(r - g))/2 - bRG = R - GBY =B - YI1 = ((r+g+b) /3)M = np.zeros((h, w))f1 = M + RG * 1jf2 = BY + I1 * 1jF1 = np.fft.fft2(f1)F2 = np.fft.fft2(f2)phaseQ1 = np.angle(F1)phaseQ2 = np.angle(F2)ifftq1 = np.fft.ifft2(np.exp(phaseQ1 * 1j))ifftq2 = np.fft.ifft2(np.exp(phaseQ2 * 1j))absq1 = np.abs(ifftq1)absq2 = np.abs(ifftq2)squareq=(absq1+absq2) * (absq1+absq2)out = cv2.GaussianBlur(squareq, (5, 5), sigma)out = cv2.normalize(out.astype('float'), None, 0, 255, cv2.NORM_MINMAX)return outdef gen_hashing(global_feature, local_feature):"""先求出列均值，在算出每一列之间的距离，最后使用序数度量来代表哈希值input:array (x,64,64)output:list (x)"""result = glcm(global_feature)y_mean = np.mean(local_feature, axis = 0)z = np.sqrt((y_mean[1:] - y_mean[:-1]) ** 2) * 1000result.extend(z)sorted_indices = sorted(range(len(result)), key=lambda i: result[i])result = [sorted_indices.index(i)+1 for i in range(len(result))]return resultdef glcm(img, levels = 32):'''https://www.cnblogs.com/xiaoliang-333/articles/16937977.htmlgraycom = greycomatrix(img, [1], [0, np.pi/4, np.pi/2, np.pi*3/4], levels=256)c = feature.greycoprops(graycom, 'contrast')  # 对比度d = feature.greycoprops(graycom, 'dissimilarity')   # 相异性h = feature.greycoprops(graycom, 'homogeneity')    # 同质性e = feature.greycoprops(graycom, 'energy')    # 能量corr = feature.greycoprops(graycom, 'correlation')    # 相关性ASM = feature.greycoprops(graycom, 'ASM')     # 角二阶矩'''from skimage.feature import graycomatrix, graycopropsimg = img.astype(np.float64)img = img * levels / 256.0img = img.astype(np.uint8)distances = [1, 1, 1, 1]  angles = [0, 45, 90, 135] #初始化一个空列表来存储GLCM矩阵统计特征glcms = []#为每个距离和角度组合计算 GLCMfor d,a in zip(distances,angles):glcm = graycomatrix(img,distances=[d],angles=[a],levels=levels,symmetric=True, normed=True)contrast = graycoprops(glcm, 'ASM')     glcms.append(contrast[0, 0])correlation = graycoprops(glcm, 'contrast')  glcms.append(correlation[0, 0])energy = graycoprops(glcm, 'correlation')    glcms.append(energy[0, 0])homogeneity = graycoprops(glcm, 'homogeneity')    glcms.append(homogeneity[0, 0])# return np.array(np.round(glcms), dtype=np.uint8)return glcmsdef dist_img(h1,h2):# distance = np.count_nonzero(np.array(list(h1)) != np.array(list(h2)))# return distance / len(h1)h1 = np.array(h1)h2 = np.array(h2)return sum(np.abs(h1-h2))/len(h1)