### Python OpenCV 中的金字塔融合原理及其实现 #### 高斯金字塔与拉普拉斯金字塔基础 在图像处理领域，高斯金字塔和拉普拉斯金字塔是两种常用的多尺度表示方法。高斯金字塔通过 `cv2.pyrDown()` 和 `cv2.pyrUp()` 函数实现向下采样和向上采样的过程[^1]。每次调用 `cv2.pyrDown()` 会将图像缩小一半，而 `cv2.pyrUp()` 则将其放大一倍。然而，在放大过程中可能会引入一些高频细节丢失的问题。 拉普拉斯金字塔则是基于高斯金字塔构建的一种差分结构，它记录了不同层次之间的差异信息。具体来说，拉普拉斯金字塔每一层由当前高斯金字塔层减去其上一层经过放大的版本构成[^3]。 #### 金字塔融合基本流程 金字塔融合是一种结合两幅或多幅图像的技术，通常应用于图像拼接、混合以及增强等领域。以下是实现该技术的主要步骤： 1. **生成高斯金字塔** 对输入的两张图片分别创建它们各自的高斯金字塔序列。 2. **生成拉普拉斯金字塔** 基于每张图片对应的高斯金字塔来计算各自拉普拉斯金字塔。 3. **组合拉普拉斯金字塔** 将两个源图像在同一位置处按照一定权重或者掩码进行合并形成新的拉普拉斯金字塔。 4. **重构最终图像** 使用新合成后的拉普拉斯金字塔逐级叠加恢复原始大小的目标图像。 下面给出一段具体的代码示例展示这一整个过程: ```python import cv2 import numpy as np def gaussian_pyramid(image, levels=6): """Generate Gaussian Pyramid""" lower = image.copy() gaussian_pyr = [lower] for _ in range(levels - 1): lower = cv2.pyrDown(lower) gaussian_pyr.append(np.float32(lower)) return gaussian_pyr def laplacian_pyramid(gaussian_pyr): """Generate Laplacian Pyramid from a given Gaussian Pyramid.""" laplacian_top = gaussian_pyr[-1] num_levels = len(gaussian_pyr) - 1 laplacian_pyr = [laplacian_top] for i in range(num_levels, 0, -1): size = (gaussian_pyr[i - 1].shape[1], gaussian_pyr[i - 1].shape[0]) gaussian_expanded = cv2.pyrUp(gaussian_pyr[i], dstsize=size) laplacian = np.subtract(gaussian_pyr[i-1], gaussian_expanded) laplacian_pyr.append(laplacian) return laplacian_pyr[::-1] def blend_images(A, B, mask, levels=6): """ Blend two images using pyramid blending. :param A: First input image :param B: Second input image :param mask: Mask defining the region to take from each image :return: Blended output image """ # Convert all inputs into float format and normalize them between [0..1]. A_float = np.float32(A) / 255.0 B_float = np.float32(B) / 255.0 mask_float = np.float32(mask) / 255.0 # Generate Gaussian pyramids for both source images along with their masks. gpA = gaussian_pyramid(A_float, levels) gpB = gaussian_pyramid(B_float, levels) gpmask = gaussian_pyramid(mask_float, levels) # Create corresponding Laplace Pyramids based on above Gaussians. lpA = laplacian_pyramid(gpA) lpB = laplacian_pyramid(gpB) # Combine these Lapsace layers according to provided mask at respective resolutions. blended_laplace = [] for la_layer, lb_layer, gm_layer in zip(lpA, lpB, reversed(gpmask)): ls = gm_layer * la_layer + (1 - gm_layer) * lb_layer blended_laplace.append(ls) # Reconstruct final combined image by collapsing this new set of Laplaces back up again. reconstructed_image = blended_laplace[0] for i in range(1, levels): size = (blended_laplace[i].shape[1], blended_laplace[i].shape[0]) expanded_reconstruction = cv2.resize(reconstructed_image, dsize=size) reconstructed_image = cv2.add(expanded_reconstruction, blended_laplace[i]) result = np.clip((reconstructed_image * 255).astype('uint8'), 0, 255) return result if __name__ == "__main__": apple = cv2.imread('./apple.jpg') orange = cv2.imread('./orange.jpg') assert(apple.shape[:2]==orange.shape[:2]), "Images must be same dimensions!" mask = np.zeros_like(apple[:, :, 0]) # Black background initially... mask[int(apple.shape[0]/2):,:] = 255 # ...with white stripe across center horizontally. out = blend_images(apple, orange, mask, levels=6) cv2.imwrite("./output_fused.png",out) ``` 此脚本实现了简单的苹果橙子水果拼接效果演示[^2][^3]。 #### 注意事项 - 输入图像需具有相同的尺寸；如果不一样，则应先调整至一致再执行后续操作。 - 掩模设计直接影响到最终结果呈现方式——可以自定义更复杂的渐变过渡区域而非简单直线分割线。 ---