o :ήc@s>ddlZddlmZmZddZddZdd Zd d d ZdS)N)_supported_float_type _to_np_modecCs,|D]}|ddkrd|}t|qdS)zEnsure all sizes in ``axis_sizes`` are odd. Parameters ---------- axis_sizes : iterable of int Raises ------ ValueError If any given axis size is even. rrzbWindow size for `threshold_sauvola` or `threshold_niblack` must not be even on any dimension. Got N) ValueError) axis_sizes axis_sizemsgr >/tmp/pip-target-vg8gfxp4/lib/python/skimage/filters/_sparse.py_validate_window_sizes r cCs8tddt|||jD}||}|dkr|S||S)zGet a view into `padded` that is offset by `idx` and scaled by `val`. If `padded` was created by padding the original image by `kernel_shape` as in correlate_sparse, then the view created here will match the size of the original image. cSs(g|]\}}}t|||d|qS))slice).0cw_sr r r !sz_get_view..r )tuplezipshape)padded kernel_shapeidxvalsl_shiftvr r r _get_views rcCsx|d|d}}t|d|jkrtdt||||}t|dd|ddD] \}}|t||||7}q,|S)aPerform correlation with a sparse kernel. Parameters ---------- image : ndarray The (prepadded) image to be correlated. kernel_shape : tuple of int The shape of the sparse filter kernel. kernel_indices : list of coordinate tuples The indices of each non-zero kernel entry. kernel_values : list of float The kernel values at each location in kernel_indices. Returns ------- out : ndarray The filtered image. Notes ----- This function only returns results for the 'valid' region of the convolution, and thus `out` will be smaller than `image` by an amount equal to the kernel size along each axis. rrz*Unexpected initial index in kernel_indicesr N)rndim RuntimeErrorrcopyr)imagerkernel_indices kernel_valuesrroutr r r _correlate_sparse)s"r%reflectc Cst|}t|j}|j|dd}|dkr|}nt|}t|jtj|dd|jD|d}t |}t ||j|dd}t t |}d|j }||vrX|g|}dg|}t ||j||S) aCompute valid cross-correlation of `padded_array` and `kernel`. This function is *fast* when `kernel` is large with many zeros. See ``scipy.ndimage.correlate`` for a description of cross-correlation. Parameters ---------- image : ndarray, dtype float, shape (M, N,[ ...,] P) The input array. If mode is 'valid', this array should already be padded, as a margin of the same shape as kernel will be stripped off. kernel : ndarray, dtype float shape (Q, R,[ ...,] S) The kernel to be correlated. Must have the same number of dimensions as `padded_array`. For high performance, it should be sparse (few nonzero entries). mode : string, optional See `scipy.ndimage.correlate` for valid modes. Additionally, mode 'valid' is accepted, in which case no padding is applied and the result is the result for the smaller image for which the kernel is entirely inside the original data. Returns ------- result : array of float, shape (M, N,[ ...,] P) The result of cross-correlating `image` with `kernel`. If mode 'valid' is used, the resulting shape is (M-Q+1, N-R+1,[ ...,] P-S+1). F)r validcSsg|] }|d|dfqS)rr )rwr r r rvsz$correlate_sparse..)moderg)npasarrayrdtypeastyperr rpadnonzerolistrrr%) r!kernelr) float_dtype padded_imagenp_modeindicesvalues corner_indexr r r correlate_sparseMs(        r8)r&) numpyr* _shared.utilsrrr rr%r8r r r r s $