图像预处理之减去RGB均值

减去RGB均值（实例以DIV2K数据集为例）

在计算机视觉领域中，一定免不了的就是图像预处理中的 逐个样本减去mean值的过程，那么为什么一定要做这一步呢？

为什么每张图片都要减去数据集均值呢？

原因：为了进行数据特征标准化，即像机器学习中的特征预处理那样对输入特征向量各维去均值再除以标准差，但由于自然图像各点像素值的范围都在0-255之间，方差大致一样，只要做去均值（减去整个图像数据集的均值或各通道关于图像数据集的均值）处理即可。

主要原理：我们默认自然图像是一类平稳的数据分布(即数据每一维的统计都服从相同分布)，此时，在每个样本上减去数据的统计平均值可以移除共同的部分，凸显个体差异。

其效果如下所示：

可以看到天空的纹理被移除了，凸显了汽车和高楼等主要特征。

最值得注意的一点是，在计算均值之前就要预先划分好训练集验证集和测试集，然后只针对训练集计算均值，否则就违背了深度学习的原则：模型训练过程仅能从训练模型中获取信息。得到训练集的均值后，对训练集验证集和测试集分别减去该均值。

以RCAN中代码为例:

class MeanShift(nn.Conv2d):
    def __init__(self, rgb_range, rgb_mean, rgb_std, sign=-1):
        super(MeanShift, self).__init__(3, 3, kernel_size=1)
        
        std = torch.Tensor(rgb_std)
        self.weight.data = torch.eye(3).view(3, 3, 1, 1) # #第一维为输出通道，第二维为输入通道
        self.weight.data.div_(std.view(3, 1, 1, 1)) # torch.eye(3).view(3, 3, 1, 1) / std.view(3, 1, 1, 1)
        self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean)
        self.bias.data.div_(std) # sign * rgb_range * torch.Tensor(rgb_mean) / std
        # (w*x+b)/std
        self.requires_grad = False

rgb_mean = (0.4488, 0.4371, 0.4040)
rgb_std = (1.0, 1.0, 1.0)
sub_mean = MeanShift(255, rgb_mean, rgb_std)
add_mean = MeanShift(255, rgb_mean, rgb_std, sign=1)
print(sub_mean)
print('sub_mean的所有属性方法', dir(sub_mean)) # dir(Object)可以获得对象的所有属性和方法
print('stride', getattr(sub_mean,'stride')) # getattr(object, name)获取对象object的属性或者方法
print('判断是否有weight属性', hasattr(sub_mean, 'weight')) # hasattr(Object, "name")可以判断对象是否有某个属性

x = imageio.imread('322878.jpg') # 载入图片
np_transpose = np.ascontiguousarray(x.transpose((2, 0, 1)))
tensor = torch.from_numpy(np_transpose).float()
tensor = tensor.unsqueeze(0)
tensor.mul_(255. / 255)
print('input',tensor, tensor.shape)
tensor = sub_mean(tensor)
print('output',tensor, tensor.shape)
tensor = add_mean(tensor)
print('reoutput', tensor, tensor.shape)

输出

MeanShift(3, 3, kernel_size=(1, 1), stride=(1, 1))
sub_mean的所有属性方法 ['__call__', '__class__', '__constants__', '__delattr__', '__dict__', '__dir__', '__doc__', '__eq__', '__format__', '__ge__', '__getattr__', '__getattribute__', '__gt__', '__hash__', '__init__', '__init_subclass__', '__le__', '__lt__', '__module__', '__ne__', '__new__', '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', '__setstate__', '__sizeof__', '__str__', '__subclasshook__', '__weakref__', '_apply', '_backend', '_backward_hooks', '_buffers', '_forward_hooks', '_forward_pre_hooks', '_get_name', '_load_from_state_dict', '_load_state_dict_pre_hooks', '_modules', '_named_members', '_parameters', '_register_load_state_dict_pre_hook', '_register_state_dict_hook', '_slow_forward', '_state_dict_hooks', '_tracing_name', '_version', 'add_module', 'apply', 'bias', 'buffers', 'children', 'cpu', 'cuda', 'dilation', 'double', 'dump_patches', 'eval', 'extra_repr', 'float', 'forward', 'groups', 'half', 'in_channels', 'kernel_size', 'load_state_dict', 'modules', 'named_buffers', 'named_children', 'named_modules', 'named_parameters', 'out_channels', 'output_padding', 'padding', 'padding_mode', 'parameters', 'register_backward_hook', 'register_buffer', 'register_forward_hook', 'register_forward_pre_hook', 'register_parameter', 'requires_grad', 'reset_parameters', 'share_memory', 'state_dict', 'stride', 'to', 'train', 'training', 'transposed', 'type', 'weight', 'zero_grad']
stride (1, 1)
判断是否有weight属性 True
input tensor([[[[201., 201., 201.,  ..., 201., 201., 201.],
          [200., 200., 199.,  ..., 200., 200., 199.],
          [200., 200., 199.,  ..., 200., 200., 200.],
          ...,
          [205., 205., 206.,  ..., 205., 205., 205.],
          [206., 206., 206.,  ..., 205., 206., 205.],
          [208., 208., 208.,  ..., 207., 207., 207.]],

         [[200., 201., 201.,  ..., 201., 201., 201.],
          [199., 199., 199.,  ..., 199., 199., 199.],
          [199., 199., 199.,  ..., 199., 199., 199.],
          ...,
          [205., 205., 205.,  ..., 205., 205., 205.],
          [205., 205., 205.,  ..., 205., 206., 205.],
          [207., 207., 207.,  ..., 207., 207., 207.]],

         [[198., 199., 199.,  ..., 199., 199., 199.],
          [197., 197., 197.,  ..., 197., 197., 197.],
          [197., 197., 197.,  ..., 197., 197., 197.],
          ...,
          [203., 203., 203.,  ..., 203., 203., 203.],
          [203., 203., 203.,  ..., 203., 204., 203.],
          [205., 205., 205.,  ..., 205., 205., 205.]]]]) torch.Size([1, 3, 1080, 1920])
output tensor([[[[ 86.5560,  86.5560,  86.5560,  ...,  86.5560,  86.5560,  86.5560],
          [ 85.5560,  85.5560,  84.5560,  ...,  85.5560,  85.5560,  84.5560],
          [ 85.5560,  85.5560,  84.5560,  ...,  85.5560,  85.5560,  85.5560],
          ...,
          [ 90.5560,  90.5560,  91.5560,  ...,  90.5560,  90.5560,  90.5560],
          [ 91.5560,  91.5560,  91.5560,  ...,  90.5560,  91.5560,  90.5560],
          [ 93.5560,  93.5560,  93.5560,  ...,  92.5560,  92.5560,  92.5560]],

         [[ 88.5395,  89.5395,  89.5395,  ...,  89.5395,  89.5395,  89.5395],
          [ 87.5395,  87.5395,  87.5395,  ...,  87.5395,  87.5395,  87.5395],
          [ 87.5395,  87.5395,  87.5395,  ...,  87.5395,  87.5395,  87.5395],
          ...,
          [ 93.5395,  93.5395,  93.5395,  ...,  93.5395,  93.5395,  93.5395],
          [ 93.5395,  93.5395,  93.5395,  ...,  93.5395,  94.5395,  93.5395],
          [ 95.5395,  95.5395,  95.5395,  ...,  95.5395,  95.5395,  95.5395]],

         [[ 94.9800,  95.9800,  95.9800,  ...,  95.9800,  95.9800,  95.9800],
          [ 93.9800,  93.9800,  93.9800,  ...,  93.9800,  93.9800,  93.9800],
          [ 93.9800,  93.9800,  93.9800,  ...,  93.9800,  93.9800,  93.9800],
          ...,
          [ 99.9800,  99.9800,  99.9800,  ...,  99.9800,  99.9800,  99.9800],
          [ 99.9800,  99.9800,  99.9800,  ...,  99.9800, 100.9800,  99.9800],
          [101.9800, 101.9800, 101.9800,  ..., 101.9800, 101.9800, 101.9800]]]],
       grad_fn=<MkldnnConvolutionBackward>) torch.Size([1, 3, 1080, 1920])
reoutput tensor([[[[201., 201., 201.,  ..., 201., 201., 201.],
          [200., 200., 199.,  ..., 200., 200., 199.],
          [200., 200., 199.,  ..., 200., 200., 200.],
          ...,
          [205., 205., 206.,  ..., 205., 205., 205.],
          [206., 206., 206.,  ..., 205., 206., 205.],
          [208., 208., 208.,  ..., 207., 207., 207.]],

         [[200., 201., 201.,  ..., 201., 201., 201.],
          [199., 199., 199.,  ..., 199., 199., 199.],
          [199., 199., 199.,  ..., 199., 199., 199.],
          ...,
          [205., 205., 205.,  ..., 205., 205., 205.],
          [205., 205., 205.,  ..., 205., 206., 205.],
          [207., 207., 207.,  ..., 207., 207., 207.]],

         [[198., 199., 199.,  ..., 199., 199., 199.],
          [197., 197., 197.,  ..., 197., 197., 197.],
          [197., 197., 197.,  ..., 197., 197., 197.],
          ...,
          [203., 203., 203.,  ..., 203., 203., 203.],
          [203., 203., 203.,  ..., 203., 204., 203.],
          [205., 205., 205.,  ..., 205., 205., 205.]]]],
       grad_fn=<MkldnnConvolutionBackward>) torch.Size([1, 3, 1080, 1920])

reference：
https://github.com/yulunzhang/RCAN
https://zhuanlan.zhihu.com/p/141644506
https://blog.csdn.net/Miss_yuki/article/details/80662017