crumpets.augmentation module

crumpets.augmentation.calc_scale_ratio(source_size, target_size, scale, scale_mode)

crumpets.augmentation.decode_image(data, color, min_height=0, min_width=0, min_factor=2)

crumpets.augmentation.decode_opencv(data, color)

crumpets.augmentation.make_transform(source_size, target_size, angle=0, scale=1, aspect=1, shift=None, hmirror=False, vmirror=False, shear=None, scale_mode='shortest', __identity__=array([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]))

crumpets.augmentation.randomize_image(im, size, background=None, color=True, angle=0, scale=1, shift=None, aspect=1, hmirror=False, vmirror=False, interp_method=1, gamma_gray=None, gamma_color=None, contrast=None, noise=None, blur=None, shear=None, is_rgb=True, scale_mode='shortest', supersampling=0, gpu_augmentation=False, do_rotate_and_resize=True)

Randomizes image according to given parameters.

Parameters

• im – image to be transformed.

• size – target size of resulting image.

• background – background color that fills areas in the output where there is no pixel data; can be number or tuple with same number of elements as channel

• color – Boolean that flags if image is black-white or colored.

• angle – degrees of rotation

• scale – Scales the image with respect to its target size. scale=1.0 scales the image to fit perfectly within the target size. Based on scale_mode either the shorter or longer edge is used as reference. scale=2.0 doubles the length of the sides, scale=0.5 halves it.

• shift – tuple of int (x,y) defining a shift of the picture, may create undefined space, if source image is moved out of target image, filled up with background color.

• aspect – float of aspect ratio change

• hmirror – boolean flag for horizontal mirror

• vmirror – boolean flag for vertical mirror

• interp_method – some interpolation method. At the moment one of: INTERP_NEAREST INTERP_LINEAR INTERP_CUBIC INTERP_LANCZOS4 INTERP_AERA

• gamma_gray – float defining a black-white gamma

• gamma_color – tuple of floats defining a rgb gamma

• contrast – float between -1 and 1 defining a contrast change

• noise – float defining a noise strength

• blur – float defining a blur intensity, i.e. the standard deviation of a gaussian filter relative to image width

• shear – float defining shear intensity, i.e. the gradient of the horizontal edges. A shear of 0.0 therefore creates a rectangular image.

• is_rgb – boolean that flags if rgb color encoding is used

• scale_mode – Either ‘shortest’ or ‘longest’. Scale the image using either shortest or longest edge as reference. ‘shortest’ crops part of the image if the aspect ratio of image and target size do not match. ‘longest’ ensures that the whole image can be fit into target size. A scale > 1.0 makes it bigger than target image, thus parts of it get cut out. A scale < 1.0 makes it smaller than target image, thus parts of the target image are undefined and filled up with background.

• supersampling – supersampling factor, 1 turns off supersampling, 2 means 4 samples per pixel, 3 means 9 samples and so on; default of 0 means choose best based on true image size, output size and scale factor

• gpu_augmentation – boolean that flags if gpu augmentations is used elsewhere and thus disables cpu augmentations in this method for all augmentations where gpu versions are available.

• do_rotate_and_resize – boolean that flags if rotation and resize operations are used. Mostly used for test cases. Should usually not be changed.

Returns

randomized image

crumpets.augmentation.rotate_and_resize(im, angle, target_size, scale, aspect, shift, method, background, hmirror, vmirror, shear=None, scale_mode='shortest', supersampling=0)