def preprocess_batch_deep_image(input_image, pixel_mean=None, pixel_std=None, rgb2bgr=True, warning=True, debug=True):
'''
this function preprocesses batch of images to a deep image,
1: from rgb to bgr
2: normalize the batch image based on mean and std
3: convert (N)HWC to NCHW
parameters:
input_image: NHWC numpy color image, where C is 3, uint8 or float32
pixel_mean: a float32 numpy array mean over 3 channels with shape of (3, ) or (1, )
pixel_std: a float32 numpy array std over 3 channels with shape of (3, ) or (1, )
rgb2bgr: true if the input image is rgb format, such that the output is bgr image
outputs:
np_image: NCHW float32 color numpy image, bgr format
'''
np_image, isnan = safe_batch_image(input_image, warning=warning, debug=debug)
if debug: assert np_image.shape[-1] == 3, 'the input image should be a color image'
if isuintnparray(np_image): np_image = np_image.astype('float32') / 255.
else: assert isfloatnparray(np_image), 'the input image-like array should be either an uint8 or float32 array'
if rgb2bgr: np_image = np_image[:, :, :, [2, 1, 0]] # from rgb to bgr, currently NHWC
# normalize the numpy image data
if pixel_mean is not None:
pixel_mean = safe_npdata(pixel_mean, warning=warning, debug=debug)
if debug:
assert pixel_mean.shape == (3, ) or pixel_mean.shape == (1, ), 'pixel mean is not correct'
assert (pixel_mean <= 1.0).all() and (pixel_mean >= 0.0).all(), 'mean value should be in range [0, 1].'
if pixel_mean.shape == (3, ): pixel_mean_reshape = np.reshape(pixel_mean, (1, 1, 1, 3))
elif pixel_mean.shape == (1, ): pixel_mean_reshape = np.reshape(np.repeat(pixel_mean, 3), (1, 1, 1, 3))
else: assert False, 'pixel mean is not correct'
np_image -= pixel_mean_reshape
if pixel_std is not None:
pixel_std = safe_npdata(pixel_std, warning=warning, debug=debug)
if debug:
assert pixel_std.shape == (3, ) or pixel_std.shape == (1, ), 'pixel std is not correct'
assert (pixel_std <= 1.0).all() and (pixel_std >= 0.0).all(), 'std value should be in range [0, 1].'
if pixel_std.shape == (3, ): pixel_std_reshape = np.reshape(pixel_std, (1, 1, 1, 3))
elif pixel_std.shape == (1, ): pixel_std_reshape = np.reshape(np.repeat(pixel_std, 3), (1, 1, 1, 3))
else: assert False, 'pixel std is not correct'
np_image /= pixel_std_reshape
np_image = np.transpose(np_image, (0, 3, 1, 2)) # NHWC to NCHW
return np_image
def unpreprocess_batch_deep_image(input_image, pixel_mean=None, pixel_std=None, bgr2rgb=True, warning=True, debug=True):
'''
this function unpreprocesses batch of deep images, which uses chw format instead of hwc format in general
1: un-normalize the batch image based on mean and std
2: convert NCHW to NHWC
3. from bgr to rgb
parameters:
input_image: NCHW / CHW float32 numpy array, where C is 3
pixel_mean: a float32 numpy array mean over 3 channels with shape of (3, ) or (1, )
pixel_std: a float32 numpy array std over 3 channels with shape of (3, ) or (1, )
bgr2rgb: true if the input image is bgr format, such that the output is rgb image
outputs:
np_image: NHWC uint8 color numpy image, rgb format
'''
np_image, isnan = safe_batch_deep_image(input_image, warning=warning, debug=debug)
if debug: assert isfloatnparray(np_image), 'the input image-like array should be either an uint8 or float32 array'
if pixel_std is not None:
pixel_std = safe_npdata(pixel_std, warning=warning, debug=debug)
if debug:
assert pixel_std.shape == (3, ) or pixel_std.shape == (1, ), 'pixel std is not correct'
assert (pixel_std <= 1.0).all() and (pixel_std >= 0.0).all(), 'std value should be in range [0, 1].'
if pixel_std.shape == (3, ): pixel_std_reshape = np.reshape(pixel_std, (1, 3, 1, 1))
elif pixel_std.shape == (1, ): pixel_std_reshape = np.reshape(np.repeat(pixel_std, 3), (1, 3, 1, 1))
else: assert False, 'pixel std is not correct'
np_image *= pixel_std_reshape
if pixel_mean is not None:
pixel_mean = safe_npdata(pixel_mean, warning=warning, debug=debug)
if debug:
assert pixel_mean.shape == (3, ) or pixel_mean.shape == (1, ), 'pixel mean is not correct'
assert (pixel_mean <= 1.0).all() and (pixel_mean >= 0.0).all(), 'mean value should be in range [0, 1].'
if pixel_mean.shape == (3, ): pixel_mean_reshape = np.reshape(pixel_mean, (1, 3, 1, 1))
elif pixel_mean.shape == (1, ): pixel_mean_reshape = np.reshape(np.repeat(pixel_mean, 3), (1, 3, 1, 1))
else: assert False, 'pixel mean is not correct'
np_image += pixel_mean_reshape
np_image = np.transpose(np_image, (0, 2, 3, 1)) # permute to [batch, height, weight, channel]
if bgr2rgb: np_image = np_image[:, :, :, [2, 1, 0]] # from bgr to rgb for color image
np_image = (np.clip(np_image, 0.0, 1.0) * 255.).astype('uint8')
return np_image