Tensorflow-API convert_image_dtype

参考：
1、https://tensorflow.google.cn/api_docs/python/tf/image/convert_image_dtype
2、http://tensorflow.org

tf.image.convert_image_dtype

将图像转换为dtype，如果需要，缩放其值。

convert_image_dtype(
    image,
    dtype,
    saturate=False,
    name=None
)

1、如果image的数值类型是int，dtype=float时会将数值缩放到[0,1)范围

#!/usr/bin/env python3
# -*- coding: UTF-8 -*-

import tensorflow as tf
from scipy import misc,ndimage


data=misc.imread('Lenna.png',mode='L') # shape 512x512
data=ndimage.zoom(data, 0.01) # 缩放为原来的0.01倍
print(data.dtype) # uint8
print(type(data)) # <class 'numpy.ndarray'>
print(data.shape) # (5, 5)
print(data)
'''
[[161 131 137 151 130]
 [ 93 181 183 153 157]
 [ 97 142 103  64 135]
 [ 46  87  84 151  46]
 [ 41  69 134 112 106]]
'''

float_image_batch = tf.image.convert_image_dtype(data, tf.float16)

sess=tf.InteractiveSession()
tf.global_variables_initializer().run()
print(float_image_batch.eval())
'''
[[ 0.63134766  0.51367188  0.53710938  0.59228516  0.50976562]
 [ 0.36474609  0.70996094  0.71777344  0.60009766  0.61572266]
 [ 0.38037109  0.55664062  0.40380859  0.25097656  0.52929688]
 [ 0.18041992  0.34106445  0.3293457   0.59228516  0.18041992]
 [ 0.1607666   0.27050781  0.52539062  0.43920898  0.41577148]]
'''

2、如果image原本数据类型为unit8,而转成unit16,dtype=float,会发现缩放后的数值存在很大的问题

#!/usr/bin/env python3
# -*- coding: UTF-8 -*-

import tensorflow as tf
from scipy import misc,ndimage
import numpy as np

data=misc.imread('Lenna.png',mode='L') # shape 512x512
data=ndimage.zoom(data, 0.01) # 缩放为原来的0.01倍
print(data.dtype) # uint8
print(type(data)) # <class 'numpy.ndarray'>
print(data.shape) # (5, 5)
data=data.astype(np.uint16)
print(data)
'''
[[161 131 137 151 130]
 [ 93 181 183 153 157]
 [ 97 142 103  64 135]
 [ 46  87  84 151  46]
 [ 41  69 134 112 106]]
'''

float_image_batch = tf.image.convert_image_dtype(data, tf.float16)

sess=tf.InteractiveSession()
tf.global_variables_initializer().run()
print(float_image_batch.eval())
'''
[[ 0.00245667  0.0019989   0.00209045  0.00230408  0.00198364]
 [ 0.00141907  0.00276184  0.00279236  0.00233459  0.00239563]
 [ 0.0014801   0.00216675  0.00157166  0.00097656  0.00205994]
 [ 0.0007019   0.00132751  0.00128174  0.00230408  0.0007019 ]
 [ 0.00062561  0.00105286  0.00204468  0.00170898  0.00161743]]
'''

3、如果image的数值类型是float，dtype=float 数值不会缩放到[0,1)范围

#!/usr/bin/env python3
# -*- coding: UTF-8 -*-

import tensorflow as tf
from scipy import misc,ndimage
import numpy as np

data=misc.imread('Lenna.png',mode='L') # shape 512x512
data=ndimage.zoom(data, 0.01) # 缩放为原来的0.01倍
print(data.dtype) # uint8
print(type(data)) # <class 'numpy.ndarray'>
print(data.shape) # (5, 5)
data=data.astype(np.float16)
print(data)
'''
[[ 161.  131.  137.  151.  130.]
 [  93.  181.  183.  153.  157.]
 [  97.  142.  103.   64.  135.]
 [  46.   87.   84.  151.   46.]
 [  41.   69.  134.  112.  106.]]
'''

float_image_batch = tf.image.convert_image_dtype(data, tf.float16)

sess=tf.InteractiveSession()
tf.global_variables_initializer().run()
print(float_image_batch.eval())
'''
[[ 161.  131.  137.  151.  130.]
 [  93.  181.  183.  153.  157.]
 [  97.  142.  103.   64.  135.]
 [  46.   87.   84.  151.   46.]
 [  41.   69.  134.  112.  106.]]
'''

总结：如果image数据类型已是float，再使用convert_image_dtype转成float类型，并不会对image的数值做归一化，则需通过别的途径进行归一化处理，如：(image-mean(image,0))/var(image,0)

并且如果没有使用数据的原本的格式，使用convert_image_dtype转成float类型会出现很大的问题

所以使用convert_image_dtype一定要检查好原数据的数据类型，否则转换会存在问题

补充：数值类型范围

符号属性	长度属性	基本型所占位数取值范围	对应numpy	对应GDAL数据类型
– – char	8	-2^7 ~ 2^7-1	np.byte/np.int8
signed – char	8	-2^7 ~ 2^7-1	np.byte/np.int8
unsigned – char	8	0 ~ 2^8-1	np.ubyte/np.uint8	GDT_Byte
[signed] short [int]	16	-2^15 ~ 2^15-1	np.int16	GDT_Int16
unsigned short [int]	16	0 ~ 2^16-1	np.uint16	GDT_UInt16
[signed] – int	32	-2^31 ~ 2^31-1	np.int32/np.int	GDT_Int32
unsigned – [int]	32	0 ~ 2^32-1	np.uint32/np.uint	GDT_UInt32
[signed] long [int]	32	-2^31 ~ 2^31-1
unsigned long [int]	32	0 ~ 2^32-1
[signed] long long [int]	64	-2^63 ~ 2^63-1	np.int64
unsigned long long [int]	64	0 ~ 2^64-1	np.uint64
– – float	32	+/- 3.40282e+038	np.flaot32/np.float	GDT_Float32
– – double	64	+/- 1.79769e+308	np.float64	GDT_Float64
– long double	96	+/- 1.79769e+308

如果图像是8bit 推荐使用 np.ubyte/np.uint8
如果图像是16bit 推荐使用 np.uint16

转成对应的float 推荐使用np.float16/np.float32