o :ήc(]@sddlmZddlZddlZddlZddlZddlm Z ddZ ddZ dd Z d d Z d;ddZddZdddZddZd?ddZdd Zd@d!d"Zd#d$Zd%d&Zd'd(Zd)d*ZdAd,d-ZdBd/d0Zd1d2Zd3d4ZdCd5d6Zd7d8Z d9d:Z!dS)D)ImageNwrapscstfdd}|S)aCreates a new function which operates on each channel Parameters ---------- single_channel_func: function Function that acts on a single color channel Returns ------- channel_func: function The same function that operates on all color channels Example ------- >>> from pymatting import * >>> import numpy as np >>> from scipy.signal import convolve2d >>> single_channel_fun = lambda x: convolve2d(x, np.ones((3, 3)), 'valid') >>> multi_channel_fun = apply_to_channels(single_channel_fun) >>> I = np.random.rand(480, 320, 3) >>> multi_channel_fun(I).shape (478, 318, 3) cstjdkrgRiSj}|d|ddtjfddtjdDdd}|t|jddt|ddS)Nrcs8g|]}dddd|fgRiqSN)copy).0c)argsimagekwargssingle_channel_func:/tmp/pip-target-vg8gfxp4/lib/python/pymatting/util/util.py +s*zAapply_to_channels..multi_channel_func..axis)lenshapereshapenpstackrangelist)r r rrresultr)r r rrmulti_channel_func"s (z-apply_to_channels..multi_channel_funcr)rrrrrapply_to_channels srcCtd||S)a0Computes the dot product of two vectors. Parameters ---------- a: numpy.ndarray First vector (if np.ndim(a) > 1 the function calculates the product for the two last axes) b: numpy.ndarray Second vector (if np.ndim(b) > 1 the function calculates the product for the two last axes) Returns ------- product: scalar Dot product of `a` and `b` Example ------- >>> import numpy as np >>> from pymatting import * >>> a = np.ones(2) >>> b = np.ones(2) >>> vec_vec_dot(a,b) 2.0 z...i,...i->...reinsumabrrr vec_vec_dot7r&cCr )aKCalculates the matrix vector product for two arrays. Parameters ---------- A: numpy.ndarray Matrix (if np.ndim(A) > 2 the function calculates the product for the two last axes) b: numpy.ndarray Vector (if np.ndim(b) > 1 the function calculates the product for the two last axes) Returns ------- product: numpy.ndarray Matrix vector product of both arrays Example ------- >>> import numpy as np >>> from pymatting import * >>> A = np.eye(2) >>> b = np.ones(2) >>> mat_vec_dot(A,b) array([1., 1.]) z...ij,...j->...ir!)Ar%rrr mat_vec_dotRr'r)cCr )aVComputes the outer product of two vectors a: numpy.ndarray First vector (if np.ndim(b) > 1 the function calculates the product for the two last axes) b: numpy.ndarray Second vector (if np.ndim(b) > 1 the function calculates the product for the two last axes) Returns ------- product: numpy.ndarray Outer product of `a` and `b` as numpy.ndarray Example ------- >>> import numpy as np >>> from pymatting import * >>> a = np.arange(1,3) >>> b = np.arange(1,3) >>> vec_vec_outer(a,b) array([[1, 2], [2, 4]]) z ...i,...jr!r#rrr vec_vec_outermsr*皙??cCsr|dks|dkr td|dks|dkrtd||kr td||k}||k}dt|}d||<d||<|S)aFixes broken trimap :math:`T` by thresholding the values .. math:: T^{\text{fixed}}_{ij}= \begin{cases} 0,&\text{if } T_{ij}<\text{lower_threshold}\\ 1,&\text{if }T_{ij}>\text{upper_threshold}\\ 0.5, &\text{otherwise}.\\ \end{cases} Parameters ---------- trimap: numpy.ndarray Possibly broken trimap lower_threshold: float Threshold used to determine background pixels, defaults to 0.1 upper_threshold: float Threshold used to determine foreground pixels, defaults to 0.9 Returns ------- fixed_trimap: numpy.ndarray Trimap having values in :math:`\{0, 0.5, 1\}` Example ------- >>> from pymatting import * >>> import numpy as np >>> trimap = np.array([0,0.1, 0.4, 0.9, 1]) >>> fix_trimap(trimap, 0.2, 0.8) array([0. , 0. , 0.5, 1. , 1. ]) rrzInvalid lower thresholdzInvalid upper thresholdz4Lower threshold must be smaller than upper thresholdg?) ValueErrorr ones_like)trimaplower_thresholdupper_thresholdis_bgis_fgfixedrrr fix_trimaps$r5cCs$zt|WdStyYdSw)aCChecks if an object is iterable Parameters ---------- obj: object Object to check Returns ------- is_iterable: bool Boolean variable indicating wether the object is iterable Example ------- >>> from pymatting import * >>> l = [] >>> isiterable(l) True TF)iter TypeError)objrrr isiterables  r9bicubiccCs^tjtjtjtjtjtjtjd}t|s#t|j |t|j |f}| ||| }|S)N)r:bilinearboxhamminglanczosnearestnone) rBICUBICBILINEARBOXHAMMINGLANCZOSNEARESTr9intwidthheightresizelower)r sizeresamplefiltersrrr_resize_pil_images rOr<cCsZt|}|dur|}|dkrdn|}||}|dur$t|||}t|d}|S)a This function can be used to load an image from a file. Parameters ---------- path: str Path of image to load. mode: str Can be "GRAY", "RGB" or something else (see PIL.convert()) Returns ------- image: numpy.ndarray Loaded image NGRAYLgo@)ropenupperconvertrOrarray)pathmoderLrMr rrr load_images   rXTcCs|jtjtjtjfvs J|jtjtjfvr#t|dddtj}|r:tj |\}}t |dkr:tj |ddt ||dS)aGiven a path, save an image there. Parameters ---------- path: str Where to save the image. image: numpy.ndarray, dtype in [np.uint8, np.float32, np.float64] Image to save. Images of float dtypes should be in range [0, 1]. Images of uint8 dtype should be in range [0, 255] make_directory: bool Whether to create the directories needed for the image path. rT)exist_okN)dtyperuint8float32float64clipastypeosrVsplitrmakedirsr fromarraysave)rVr make_directory directory_rrr save_image s ricCst|jdvs J|jtjtjtjfvsJ|jtjtjfvr,t|dddtj}t|jdkr=tj |gdddS|jddkrNtj |gdddS|jddkrW|S|jddkrk|d d d d d dfSt d |j) aConvertes an image to rgb8 color space Parameters ---------- image: numpy.ndarray Image to convert Returns ------- image: numpy.ndarray Converted image with same height and width as input image but with three color channels Example ------- >>> from pymatting import * >>> import numpy as np >>> I = np.eye(2) >>> to_rgb8(I) array([[[255, 255, 255], [ 0, 0, 0]], [[ 0, 0, 0], [255, 255, 255]]], dtype=uint8) )rrYrrrjrrNzInvalid image shape:) rrr[rr\r]r^r_r`r concatenater-)r rrrto_rgb8$s rmc Cs|D]}|dur|jtjtjfvsJqt|}|dkrdS|dur9|dur9ttt|}||d|}n|durF||d|}n |durR||d|}dd|D}tdd|D}tdd|D}td d|Ddd } | dkrt |D]K\} }|durt|j d kr|ddddtj f}|j d dkrtj |g| d d }|j d d kr| dkrt |tj|j dd |jd}||| <q~|durtdd|D}tj||||| f|d} t|D]I} t|D]B} | | |} | t|krn3|| }|dur1||j d|j dd}|| | || ||j d| || ||j df<qq| j d dkrF| dddddf} | S)aUPlots a grid of images. Parameters ---------- images : list of numpy.ndarray List of images to plot nx: int Number of rows ny: int Number of columns dtype: type Data type of output array Returns ------- grid: numpy.ndarray Grid of images with datatype `dtype` NrrcSsg|] }|dur|jqSrrr r rrrrvszmake_grid..cs|]}|dVqdS)rNrr rrrr xzmake_grid..csrp)rNrrqrrrrryrscSs g|] }t|dkr|dqS)r)rrqrrrrzs )defaultrrrjrkr[css|] }|dur|jVqdSrrurorrrrrsr)r[rr]r^rrGceilsqrtmax enumeraternewaxisrl stack_imagesonesnextzerosrr)imagesnxnyr[r nshapeshwdiryxrrr make_gridPsd     2rcCs8t|}t|dddtj}t|}|dS)zPlot grid of images. Parameters ---------- images : list of numpy.ndarray List of images to plot height : int, matrix Height in pixels the output grid, defaults to 512 rYrN)rrr_r`r\rrdshow)rgridrrr show_imagess   rc Cs|r|}|}|}|dkrtjd|dd|dkr(tjd|dd|jtjtjfvr;tjd|jdd||k}||k}| dkrOt d || dkr[t d |||B}|} |||| fS) aThis function splits the trimap into foreground pixels, background pixels, and unknown pixels. Foreground pixels are pixels where the trimap has values larger than or equal to `fg_threshold` (default: 0.9). Background pixels are pixels where the trimap has values smaller than or equal to `bg_threshold` (default: 0.1). Pixels with other values are assumed to be unknown. Parameters ---------- trimap: numpy.ndarray Trimap with shape :math:`h \times w` flatten: bool If true np.flatten is called on the trimap Returns ------- is_fg: numpy.ndarray Boolean array indicating which pixel belongs to the foreground is_bg: numpy.ndarray Boolean array indicating which pixel belongs to the background is_known: numpy.ndarray Boolean array indicating which pixel is known is_unknown: numpy.ndarray Boolean array indicating which pixel is unknown bg_threshold: float Pixels with smaller trimap values will be considered background. fg_threshold: float Pixels with larger trimap values will be considered foreground. Example ------- >>> import numpy as np >>> from pymatting import * >>> trimap = np.array([[1,0],[0.5,0.2]]) >>> is_fg, is_bg, is_known, is_unknown = trimap_split(trimap) >>> is_fg array([ True, False, False, False]) >>> is_bg array([False, True, False, False]) >>> is_known array([ True, True, False, False]) >>> is_unknown array([False, False, True, True]) z:Trimap values should be in [0, 1], but trimap.min() is %s.rj stacklevel?z:Trimap values should be in [0, 1], but trimap.max() is %s.zfUnexpected trimap.dtype %s. Are you sure that you do not want to use np.float32 or np.float64 instead?rz7Trimap did not contain background values (values <= %f)z7Trimap did not contain foreground values (values >= %f)) flattenminrxwarningswarnr[rr]r^sumr-) r/r bg_threshold fg_threshold min_value max_valuer3r2is_known is_unknownrrr trimap_splitsD-   rcCst|jdks|jddkrtjdt|jdd|}|}|dkr/tjd|dd|dkr>> import numpy as np >>> from pymatting import check_image >>> image = (np.random.randn(64, 64, 2) * 255).astype(np.int32) >>> sanity_check_image(image) __main__:1: UserWarning: Expected RGB image of shape (?, ?, 3), but image.shape is (64, 64, 2). __main__:1: UserWarning: Image values should be in [0, 1], but image.min() is -933. __main__:1: UserWarning: Image values should be in [0, 1], but image.max() is 999. __main__:1: UserWarning: Unexpected image.dtype int32. Are you sure that you do not want to use np.float32 or np.float64 instead? rjrz=Expected RGB image of shape (?, ?, 3), but image.shape is %s.rrz8Image values should be in [0, 1], but image.min() is %s.rz8Image values should be in [0, 1], but image.max() is %s.zeUnexpected image.dtype %s. Are you sure that you do not want to use np.float32 or np.float64 instead?N) rrrrstrrrxr[rr]r^)r rrrrrsanity_check_images6 rcCs:t|jdkr|ddddtjf}||d||S)aThis function composes a new image for given foreground image, background image and alpha matte. This is done by applying the composition equation .. math:: I = \alpha F + (1-\alpha)B. Parameters ---------- foreground: numpy.ndarray Foreground image background: numpy.ndarray Background image alpha: numpy.ndarray Alpha matte Returns ------- image: numpy.ndarray Composed image as numpy.ndarray Example ------- >>> from pymatting import * >>> foreground = load_image("data/lemur/lemur_foreground.png", "RGB") >>> background = load_image("data/lemur/beach.png", "RGB") >>> alpha = load_image("data/lemur/lemur_alpha.png", "GRAY") >>> I = blend(foreground, background, alpha) rNrrrrrz) foreground backgroundalpharrrblendGsrcGsdd|D}tj|ddS)a This function stacks images along the third axis. This is useful for combining e.g. rgb color channels or color and alpha channels. Parameters ---------- *images: numpy.ndarray Images to be stacked. Returns ------- image: numpy.ndarray Stacked images as numpy.ndarray Example ------- >>> from pymatting.util.util import stack_images >>> import numpy as np >>> I = stack_images(np.random.rand(4,5,3), np.random.rand(4,5,3)) >>> I.shape (4, 5, 6) cSs6g|]}t|jdkr |n |ddddtjfqS)rjNrrorrrrs(z stack_images..rr)rrl)rrrrr{ksr{cCs|t|jd|j}|S)aCalculate the sum of each row of a matrix Parameters ---------- A: np.ndarray or scipy.sparse.spmatrix Matrix to sum rows of Returns ------- row_sums: np.ndarray Vector of summed rows Example ------- >>> from pymatting import * >>> import numpy as np >>> A = np.random.rand(2,2) >>> A array([[0.62750946, 0.12917617], [0.8599449 , 0.5777254 ]]) >>> row_sum(A) array([0.75668563, 1.4376703 ]) r)dotrr|rr[)r(row_sumsrrrrow_sumsrrcCs6t|}d|||k<d|}tj|}||}|S)a Normalize the rows of a matrix Rows with sum below threshold are left as-is. Parameters ---------- A: scipy.sparse.spmatrix Matrix to normalize threshold: float Threshold to avoid division by zero Returns ------- A: scipy.sparse.spmatrix Matrix with normalized rows Example ------- >>> from pymatting import * >>> import numpy as np >>> A = np.arange(4).reshape(2,2) >>> normalize_rows(A) array([[0. , 1. ], [0.4, 0.6]]) r)rscipysparsediagsr)r( thresholdrrow_normalization_factorsDrrrnormalize_rowss    rFcCs\|rtt||}tt||}||fSt|}t|}t||\}}||fS)apCalculates image pixel coordinates for an image with a specified shape Parameters ---------- width: int Width of the input image height: int Height of the input image flatten: bool Whether the array containing the coordinates should be flattened or not, defaults to False Returns ------- x: numpy.ndarray x coordinates y: numpy.ndarray y coordinates Example ------- >>> from pymatting import * >>> x, y = grid_coordinates(2,2) >>> x array([[0, 1], [0, 1]]) >>> y array([[0, 0], [1, 1]]) )rtilearangerepeatmeshgrid)rHrIrrrrrrgrid_coordinatess rcCst|}t|}||}tj||tjd}tj||tjd} tj||tjd} d} t||dd\} } t|||D]?\}}}t | |d|d}t | |d|d}| | ||| | |<|||| | | |<|| | | |<| |7} q>t j j | || ff||fd}|S)aCalculates a convolution matrix that can be applied to a vectorized image Additionaly, this function allows to specify which pixels should be used for the convoltion, i.e. .. math:: \left(I * K\right)_{ij} = \sum_k K_k I_{i+{\Delta_y}_k,j+{\Delta_y}_k}, where :math:`K` is the flattened convolution kernel. Parameters ---------- width: int Width of the input image height: int Height of the input image kernel: numpy.ndarray Convolutional kernel dx: numpy.ndarray Offset in x direction dy: nunpy.ndarray Offset in y direction Returns ------- M: scipy.sparse.csr_matrix Convolution matrix rurTrrrn) rasarrayrrr~int32r^rzipr_rr csr_matrix)rHrIkerneldxdyweightscountri_indsj_indsvalueskrrdx2dy2weightx2y2r(rrrsparse_conv_matrix_with_offsetss" rcCsD|j\}}t||dd\}}||d8}||d8}t|||||S)aRCalculates a convolution matrix that can be applied to a vectorized image Parameters ---------- width: int Width of the input image height: int Height of the input image kernel: numpy.ndarray Convolutional kernel Returns ------- M: scipy.sparse.csr_matrix Convolution matrix Example ------- >>> from pymatting import * >>> import numpy as np >>> sparse_conv_matrix(3,3,np.ones((3,3))) <9x9 sparse matrix of type '' with 49 stored elements in Compressed Sparse Row format> Trr)rrr)rHrIrkhkwrrrrrsparse_conv_matrix)s   rcCs0|rt|}|t|}tj|}||}|S)aCalculates the random walk normlized Laplacian matrix from the weight matrix Parameters ---------- W: numpy.ndarray Array of weights normalize: bool Whether the rows of W should be normalized to 1, defaults to True regularization: float Regularization strength, defaults to 0, i.e. no regularizaion Returns ------- L: scipy.sparse.spmatrix Laplacian matrix Example ------- >>> from pymatting import * >>> import numpy as np >>> weights_to_laplacian(np.ones((4,4))) matrix([[ 0.75, -0.25, -0.25, -0.25], [-0.25, 0.75, -0.25, -0.25], [-0.25, -0.25, 0.75, -0.25], [-0.25, -0.25, -0.25, 0.75]]) )rrrrr)W normalizeregularizationrrrQrrrweights_to_laplacianJs   rcCs*t|}|}|}||||S)aNormalizes an array such that all values are between 0 and 1 Parameters ---------- values: numpy.ndarray Array to normalize Returns ------- result: numpy.ndarray Normalized array Example ------- >>> from pymatting import * >>> import numpy as np >>> normalize(np.array([0, 1, 3, 10])) array([0. , 0.1, 0.3, 1. ]) )rrrrx)rr$r%rrrrps rcCs||d|S)a1Divides a number x by another integer n and rounds up the result Parameters ---------- x: int Numerator n: int Denominator Returns ------- result: int Result Example ------- >>> from pymatting import * >>> div_round_up(3,2) 2 rr)rrrrr div_round_upsr)r+r,)r:)NNr<)T)NNN)Tr+r,)r)F)Tr)"PILrnumpyr scipy.sparserrar functoolsrrr&r)r*r5r9rOrXrirmrrrrrr{rrrrrrrrrrrrs<  . 5    ,W Y5$  (*2 !&