RGB降噪自编码器输出纯白图像问题求助

问题概述 #

实现降噪自编码器处理CIFAR-100带噪声图像时，预测输出全为纯白图像，相关代码及输出结果如下：

原模型层代码 #

input_img = keras.Input(shape=(32, 32, 3))
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(input_img)
x = layers.MaxPooling2D((2, 2), padding='same')(x)
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)
x = layers.MaxPooling2D((2, 2), padding='same')(x)
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)
encoded = layers.MaxPooling2D((2, 2), padding='same')(x)
encoded = x  # 此处错误覆盖了池化后的编码特征
x = layers.Conv2DTranspose(32, (3, 3), strides=(2, 2), activation='relu', padding='same')(encoded)
x = layers.Conv2DTranspose(32, (3, 3), strides=(2, 2), activation='relu', padding='same')(x)
decoded = layers.Conv2D(3, (3, 3), activation='relu', padding='same')(x)  # 激活函数错误
autoencoder = keras.Model(input_img, decoded)
s = 20 * 5000
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(0.01, s, 0.1)
autoencoder.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=lr_schedule), loss='mean_squared_error',metrics=['accuracy'])
autoencoder.summary()

原数据加载代码 #

(x_train, _), (x_test, _) = cifar100.load_data()

x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
x_train = np.reshape(x_train, (len(x_train), 32, 32, 3))
x_test = np.reshape(x_test, (len(x_test), 32, 32, 3))
x_train = x_train[:5000]
x_test = x_test[6000:7000]

noise_factor = 0.2
x_train_noisy = x_train + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=x_train.shape)
x_test_noisy = x_test + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=x_test.shape)
x_train_noisy = np.clip(x_train_noisy, 0., 1.)
x_test_noisy = np.clip(x_test_noisy, 0., 1.)

原训练与预测代码 #

autoencoder.fit(x_train_noisy, x_train,
                epochs=50,
                batch_size=128,
                shuffle=True)
decoded_imgs = autoencoder.predict(x_test_noisy)

for i in range(5):
    # Display original
    ax = plt.subplot(2, 5, i + 1)
    plt.imshow(x_test_noisy[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
    # Display reconstruction
    ax = plt.subplot(2, 5, i + 1 + 5)
    plt.imshow(decoded_imgs[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

输出结果 #

问题根源与修复方案 #

1. 编码器特征覆盖错误 #

原代码中encoded = x直接覆盖了MaxPooling2D后的编码特征，导致编码器未完成下采样，输入维度未被压缩，后续解码器上采样后维度不匹配。删除encoded = x这一行，保留池化后的编码结果。

2. 解码器输出激活函数错误 #

最后一层Conv2D使用relu激活，输出范围无上限，超出图像数据的[0,1]区间，导致plt.imshow将超过1的像素显示为纯白。将激活函数改为sigmoid，把输出限制在[0,1]范围内。

3. 解码器维度缺失上采样 #

原编码器经过三次MaxPooling2D后，输入32x32会变为4x4（32→16→8→4），但解码器只有两次Conv2DTranspose（步长2），只能还原到16x16，与输入32x32维度不匹配。增加一次Conv2DTranspose上采样，将维度还原为32x32。

4. 优化器与学习率调整 #

初始学习率0.01搭配SGD收敛慢且易不稳定，改用Adam优化器（学习率0.001）更适合自编码器任务，收敛速度和稳定性更好。

修改后的完整代码 #

修复后的模型层代码 #

input_img = keras.Input(shape=(32, 32, 3))
# 编码器部分
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(input_img)
x = layers.MaxPooling2D((2, 2), padding='same')(x)  # 32→16
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)
x = layers.MaxPooling2D((2, 2), padding='same')(x)  #16→8
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)
x = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)
encoded = layers.MaxPooling2D((2, 2), padding='same')(x)  #8→4，保留此编码结果

# 解码器部分
x = layers.Conv2DTranspose(32, (3, 3), strides=(2, 2), activation='relu', padding='same')(encoded)  #4→8
x = layers.Conv2DTranspose(32, (3, 3), strides=(2, 2), activation='relu', padding='same')(x)  #8→16
x = layers.Conv2DTranspose(32, (3, 3), strides=(2, 2), activation='relu', padding='same')(x)  #16→32，新增上采样
decoded = layers.Conv2D(3, (3, 3), activation='sigmoid', padding='same')(x)  # 改用sigmoid激活

autoencoder = keras.Model(input_img, decoded)
# 改用Adam优化器，学习率0.001
autoencoder.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), 
                    loss='mean_squared_error',
                    metrics=['accuracy'])
autoencoder.summary()

训练与预测代码（无需修改） #

autoencoder.fit(x_train_noisy, x_train,
                epochs=50,
                batch_size=128,
                shuffle=True)
decoded_imgs = autoencoder.predict(x_test_noisy)

for i in range(5):
    # 显示带噪声原图
    ax = plt.subplot(2, 5, i + 1)
    plt.imshow(x_test_noisy[i])
    plt.axis('off')
    # 显示修复后的图像
    ax = plt.subplot(2, 5, i + 1 + 5)
    plt.imshow(decoded_imgs[i])
    plt.axis('off')
plt.show()

效果验证 #

修改后重新训练，解码器输出维度与输入一致（32x32x3），输出像素值被限制在[0,1]区间，预测结果将正常还原带噪声图像，不会出现纯白情况。

内容的提问来源于stack exchange，提问作者Mohamed Elsayed