preprocessing.py 文件源码

def compute_dct_features(X, image_shape, no_coeff=30, method='zigzag'):
    """
    compute 2D-dct features of a given image.
    Type 2 DCT and finds the DCT coefficents with the largest mean normalized variance
    :param X: 1 dimensional input image in 'c' format
    :param image_shape: image shape
    :param no_coeff: number of coefficients to extract
    :param method: method to extract coefficents, zigzag, variance
    :return: dct features
    """
    X_dct = fft.dct(X, norm='ortho')

    if method == 'zigzag':
        out = np.zeros((len(X_dct), no_coeff), dtype=X_dct.dtype)
        for i in xrange(len(X_dct)):
            image = X_dct[i].reshape(image_shape)
            out[i] = zigzag(image)[1:no_coeff + 1]
        return out
    elif method == 'rel_variance':
        X_dct = X_dct[:, 1:]
        # mean coefficient per frequency
        mean_dct = np.mean(X_dct, 0)
        # mean normalize
        mean_norm_dct = X_dct - mean_dct
        # find standard deviation for each frequency component
        std_dct = np.std(mean_norm_dct, 0)
        # sort by largest variance
        idxs = np.argsort(std_dct)[::-1][:no_coeff]
        # return DCT coefficients with the largest variance
        return X_dct[:, idxs]
    elif method == 'variance':
        X_dct = X_dct[:, 1:]
        # find standard deviation for each frequency component
        std_dct = np.std(X_dct, 0)
        # sort by largest variance
        idxs = np.argsort(std_dct)[::-1][:no_coeff]
        # return DCT coefficients with the largest variance
        return X_dct[:, idxs]
    elif method == 'energy':
        X_dct = X_dct[:, 1:]
        X_sum = np.abs(X_dct)
        X_sum = np.sum(X_sum, 0)
        idxs = np.argsort(X_sum)[::-1][:no_coeff]
        return X_dct[:, idxs]
    else:
        raise NotImplementedError("method not implemented, use only 'zigzag', 'variance', 'rel_variance")