python - 如何在Python和Keras中通过深度学习进行光照估计后显示颜色校正图像？

Question

我正在使用 Python 和 Keras 进行深度学习的图像色彩校正的监督方法。我完成了数据训练并预测了测试数据的值。现在，我想显示数据集中的一些颜色校正图像，以便我可以将它们与原始图像进行视觉比较。我陷入了循环，我不知道如何解决这。有人可以帮助我编写代码或一些提示吗？

我已经比较了预测照明和地面真实照明的数量，但我想绘制它们在预测照明下的外观图片。我正在使用三倍交叉验证，这使得它变得更加困难。我有 1000 多张图像，但为了简单起见，我们假设只有 12 张。我将训练其中 8 个并测试其中 4 个。

#this is the part where the training and testing is happening, images are loaded
#in xs variable and ground truth illumination is loaded in ys variable

for i in range (3):
    print('\nFold ',i)
    X_train = xs [folds[i]==0, :]
    X_test = xs [folds[i]==1, :]
    Y_train = ys [folds[i]==0, :]
    Y_test = np.zeros((4,3), dtype=np.uint8)

    model = None
    model = create_model()
    history = model.fit(X_train, Y_train, epochs=10, batch_size=8)

    Y_test = model.predict(X_test, batch_size=4)
    print("Predicted values for fold %d:" % i, Y_test)
    for y in Y_test[:]:
        predicted.append(y)

这部分代码运行完美，我不知道的是如何在使用预测照明进行颜色校正后绘制甚至保存这 12 个图像中的每一个。

编辑:我已经提取了每张照片的预测值。如何将它们应用到图像上？

Answer 1

如果我理解正确的话，您希望使用模型预测的光源对偏色图像进行白平衡。您的预测由 3 个值(比方说 [alpha、beta、ceta])组成，它们是将应用于色偏图像的每个 channel (蓝色、绿色、红色)的校正增益。

但是，在应用校正增益之前，您需要对图像执行 Gamma 线性化(更多信息请参见here)。

这里有一些示例代码可以帮助您:

import cv2
import numpy as np

def gamma_decode(B_gamma, G_gamma, R_gamma):
    B_gamma = B_gamma/255
    G_gamma = G_gamma/255
    R_gamma = R_gamma/255 

    gamma = 1/2.2
    B_gamma_decode = 255*(B_gamma**(1/gamma)) 
    G_gamma_decode = 255*(G_gamma**(1/gamma))
    R_gamma_decode = 255*(R_gamma**(1/gamma))
    return (B_gamma_decode, G_gamma_decode, R_gamma_decode)


def gamma_encode(B_channel, G_channel, R_channel):
    B_channel = B_channel/255
    G_channel = G_channel/255
    R_channel = R_channel/255

    gamma = 1/2.2
    if np.all(B_channel <= 0):
        B_gamma_cor = (B_channel**(gamma + 0j))
        B_gamma_cor = 255*(abs(B_gamma_cor))
    else:
        B_gamma_cor = 255*(B_channel**gamma)

    if np.all(G_channel <= 0):
        G_gamma_cor = (G_channel**(gamma + 0j))
        G_gamma_cor = 255*(abs(G_gamma_cor))
    else:
        G_gamma_cor = 255*(G_channel**gamma)

    if np.all(R_channel <= 0):
        R_gamma_cor = (R_channel**(gamma + 0j))
        R_gamma_cor = 255*(abs(R_gamma_cor))
    else:
        R_gamma_cor = 255*(R_channel**gamma)

    return (B_gamma_cor, G_gamma_cor, R_gamma_cor)


def white_balance(img, pred_illum) 
   B_channel, G_channel, R_channel = cv2.split(img)
   alpha, beta, ceta = pred_illum

   #Gamma_decoding
   B_channel, G_channel, R_channel = gamma_decode(B_channel, G_channel, R_channel)

   #Correction
   B_cor = (alpha*B_channel)
   G_cor = (beta*G_channel)
   R_cor = (ceta*R_channel)

   #Gamma encoding
   B_cor, G_cor, R_cor = gamma_encode(B_cor, G_cor, R_cor)

   #Convert to uint8 to display
   B_cor = B_cor.astype(np.uint8)
   G_cor = G_cor.astype(np.uint8)
   R_cor = R_cor.astype(np.uint8)
   img_white_balanced = cv2.merge((B_cor, G_cor, R_cor))
   return img_white_balanced