对图像应用离散余弦变换(DCT)时生成全黑图像，求问题排查

我编写了一段Python代码用于对图像执行离散余弦变换(DCT)：加载灰度图像并调整为224x224尺寸，执行DCT后进行归一化并转换为3通道RGB图像显示。但运行后得到全黑图像，无法定位代码错误，附上效果截图与代码：

代码如下： #

import cv2
import numpy as np
import matplotlib.pyplot as plt

def perform_dct(image_path):
    # Load the image in grayscale
    img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    if img is None:
        raise ValueError(f"Image at {image_path} could not be loaded.")
    # Resize the image to 224x224 (size expected by ViT model)
    img = cv2.resize(img, (224, 224))
    # Perform DCT
    dct_img = cv2.dct(np.float32(img))

    # Normalize the DCT image to the range [0, 1]
    dct_img_min = np.min(dct_img)
    dct_img_max = np.max(dct_img)
    dct_img_normalized = (dct_img - dct_img_min) / (dct_img_max - dct_img_min)
    
    # Convert to 3-channel RGB image
    dct_img_normalized = (dct_img_normalized * 255).astype(np.uint8)
    dct_img_rgb = cv2.merge([dct_img_normalized, dct_img_normalized, dct_img_normalized])

    plt.imshow(dct_img_rgb)
    plt.title("DCT of Imagesss")
    plt.show()

    return dct_img_rgb

效果截图： #

问题排查与修复方案 #

错误原因 #

1. DCT输入范围错误：OpenCV的cv2.dct()要求输入是归一化到[0,1]的浮点型图像，你直接将0-255的uint8图像转成float32就传入，会导致DCT结果的动态范围极大——直流分量（左上角）数值极高，其余高频分量数值极小，归一化后几乎所有像素都趋近于0，显示为全黑。
2. 整幅图像DCT的可视化缺陷：直接对224x224整图做DCT，高频分量的能量占比极低，即使输入正确，可视化效果也会偏暗、细节模糊，常规做法是对图像分块（比如8x8）执行DCT。

修正后的代码 #

import cv2
import numpy as np
import matplotlib.pyplot as plt

def perform_dct(image_path):
    # Load the image in grayscale
    img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    if img is None:
        raise ValueError(f"Image at {image_path} could not be loaded.")
    # Resize the image to 224x224
    img = cv2.resize(img, (224, 224))
    
    # 关键修正1：先将图像归一化到[0,1]的浮点型
    img_float = img / 255.0
    
    # 可选优化：分块执行DCT（8x8块，符合JPEG等标准做法）
    block_size = 8
    dct_img = np.zeros_like(img_float)
    for i in range(0, img_float.shape[0], block_size):
        for j in range(0, img_float.shape[1], block_size):
            block = img_float[i:i+block_size, j:j+block_size]
            dct_img[i:i+block_size, j:j+block_size] = cv2.dct(block)
    
    # 归一化到[0,1]
    dct_img_min = np.min(dct_img)
    dct_img_max = np.max(dct_img)
    dct_img_normalized = (dct_img - dct_img_min) / (dct_img_max - dct_img_min)
    
    # 转3通道RGB并显示
    dct_img_normalized = (dct_img_normalized * 255).astype(np.uint8)
    dct_img_rgb = cv2.merge([dct_img_normalized, dct_img_normalized, dct_img_normalized])

    plt.imshow(dct_img_rgb)
    plt.title("DCT of Image")
    plt.show()

    return dct_img_rgb

修正说明 #

• 先将原始灰度图像除以255，转成[0,1]范围的float32数据，符合cv2.dct()的输入要求，避免动态范围失衡。
• 加入分块DCT处理后，可视化结果会呈现出清晰的块状DCT系数分布，左上角的低频分量更突出，高频分量也能看到细节。
• 后续的归一化和转RGB逻辑保持不变，现在能正确映射到0-255的灰度范围，显示正常。

内容的提问来源于stack exchange，提问作者Mohammad Rizabul