Python实现不同视差图计算算法

# -*- coding: utf-8 -*- """ Created on Sun Mar 3 16:49:56 2024 @author: 76003 """ import numpy as np from scipy.sparse import lil_matrix, csr_matrix from scipy.sparse.csgraph import dijkstra import cv2 # 读取图像 left_img = cv2.imread('left.png', 0) right_img = cv2.imread('right.png', 0) # 计算灰度差值矩阵 diff_map = np.abs(left_img.astype(np.float32) - right_img.astype(np.float32)) # 数据项 data_term = diff_map ** 2 # 先验项（默认为0） prior_term = np.zeros_like(data_term) # 定义节点 num_pixels = len(diff_map.flat) # 定义边 edges = [(i, j) for i in range(num_pixels) for j in range(i + 1, num_pixels)] # 定义边上的权重 weights = data_term.ravel()[edges] # 创建稀疏邻接矩阵 A = lil_matrix((num_pixels, num_pixels), dtype=np.float32) for i, j in edges: A[i, j] = weights[j] A[j, i] = weights[j] # 转换为压缩稀疏列矩阵 A = csr_matrix(A) import numpy as np from scipy.ndimage import gaussian_filter from skimage.measure import compare_ssim from skimage.feature import peak_local_max disp_range = 16 # 视差范围 num_disps = disp_range + 1 window_size = 9 # 窗口大小 sigma_cost = 3 # 成本函数的高斯滤波器的标准偏差 sigma_smooth = 8 # 平滑迭代的标准偏差 lambda_reg = 0.1 # 正则化因子 min_grad = 3 # 最小梯度阈值 max_iters = 10 # 最大迭代次数 def create_initial_disparity(image_left, image_right): """ 创建初始视差图 """ height, width = image_left.shape[:2] return np.arange(width).reshape(1, -1)[:, np.newaxis] def compute_cost_volume(image_left, image_right, initial_disp): """ 计算成本体 """ cost_volume = np.zeros((height, width, num_disps)) for i in range(num_disps): disp = initial_disp[i] cost_volume[:,:,i] = compare_ssim(image_left[:,disp:], image_right[:,:-disp], multichannel=True) return cost_volume def aggregate_costs(cost_volume, sigma_cost): """ 聚合成本 """ return gaussian_filter(cost_volume, sigma=sigma_cost, mode='nearest') def smooth Disp (disparity_map, sigma_smooth): """ 平滑视差图 """ return gaussian_filter(disparity_map, sigma=sigma_smooth, mode='nearest') def optimize_disparity(cost_volume, initial_disp, lambda_reg, min_grad, max_iters): """ 优化视差 """ disparity = initial_disp.copy() for _ in range(max_iters): gradient = np.gradient(disparity) smoothness_term = np.abs(gradient.sum(axis=(0, 1))) data_term = cost_volume[disparity.astype(int), :, :] new_disparity = (data_term / (smoothness_term + lambda_reg)).argmax(axis=-1) if np.max(np.abs(new_disparity-disp)) < min_grad: break disparity = new_disparity return disparity def gc_stereo(image_left, image_right): initial_disp = create_initial_disparity(image_left, image_right) cost_volume = compute_cost_volume(image_left, image_right, initial_disp) aggregated_cost_volume = aggregate_costs(cost_volume, sigma_cost) optimized_disp = optimize_disparity(aggregated_cost_volume, initial_disp, lambda_reg, min_grad, max_iters) return optimized_disp plt.imshow(optimized_disp,"gray") plt.show()