使用快速卷积在python中进行人脸检测

Question

我必须使用快速卷积方法对基本面部检测进行编程以计算相关性。到目前为止，我有这段代码：

def template_matching_fast_convolution(face_data, template_data):
    #convert to greyscale
    face_grey = color.rgb2grey(face_data)
    template_grey = color.rgb2grey(template_data)

    #apply zero padding to get pictures of the same size
    shape_face = face_grey.shape
    shape_template = template_grey.shape
    rows_face = shape_face[0]
    columns_face = shape_face[1]
    rows_template = shape_template[0]
    columns_template = shape_template[1]

    extra_pad_row_face = 0
    extra_pad_row_template = 0
    extra_pad_columns_face = 0
    extra_pad_colummns_template = 0

    if rows_face % 2 == 1:
        extra_pad_row_face = 1
    if rows_template % 2 == 1:
        extra_pad_row_template = 1
    if columns_face % 2 == 1:
        extra_pad_columns_face = 1
    if columns_template % 2 == 1:
        extra_pad_colummns_template = 1

    if rows_face < rows_template:
        rows_diff_div_2 = int((rows_template - rows_face) / 2)
        face_grey = np.pad(face_grey, ((rows_diff_div_2,rows_diff_div_2 + extra_pad_row_face), (0, 0)), 'constant', constant_values=(0))
    elif rows_template < rows_face:
        rows_diff_div_2 = int((rows_face - rows_template) / 2)
        template_grey = np.pad(template_grey, ((rows_diff_div_2,rows_diff_div_2 + extra_pad_row_template), (0, 0)), 'constant', constant_values=(0))

    if columns_face < columns_template:
        columns_diff_div_2 = int((columns_template - columns_face) / 2)
        face_grey = np.pad(face_grey, ((0, 0), (columns_diff_div_2,columns_diff_div_2 + extra_pad_columns_face)), 'constant', constant_values=(0))
    elif columns_template < columns_face:
        columns_diff_div_2 = int((columns_face - columns_template) / 2)
        template_grey = np.pad(template_grey, ((0, 0), (columns_diff_div_2,columns_diff_div_2 + extra_pad_colummns_template)), 'constant', constant_values=(0))

    #extra padding to make it square

    shape_face = face_grey.shape
    shape_template = template_grey.shape
    rows_face = shape_face[0]
    columns_face = shape_face[1]
    rows_template = shape_template[0]
    columns_template = shape_template[1]

    max_dimension = max(rows_face, rows_template, columns_template, columns_face)

    square_pad_rows_face = int((max_dimension - rows_face) / 2)
    square_pad_columns_face = int((max_dimension - columns_face) / 2)
    square_pad_rows_template = int((max_dimension - rows_template) / 2)
    square_pad_columns_template = int((max_dimension - columns_template) / 2)

    face_grey = np.pad(face_grey, ((square_pad_rows_face,square_pad_rows_face), (square_pad_columns_face, square_pad_columns_face)), 'constant', constant_values=(0))
    template_grey = np.pad(template_grey, ((square_pad_rows_template,square_pad_rows_template), (square_pad_columns_template, square_pad_columns_template)), 'constant', constant_values=(0))

    #apply 2D-fft
    face_fft = np.fft.fft2(face_grey)
    template_fft = np.fft.fft2(template_grey)

    #multipy the spectres in the frequency domain
    multiplied_spectrum = np.fft.fftshift(np.dot(face_fft, template_fft)).real

    #calculate the inverse fft of the spectrum
    ifft_spectrum = np.fft.ifft2(multiplied_spectrum).real

    return ifft_spectrum

face_data是一个人的照片，而template_data是一张普通人的照片（预计算）。

我不确定我是否正确地接近这一点。我是否还应根据脸部图片的大小调整模板大小？另外，如何从返回的反向光谱中确定面部？我知道幅度最大的地方是相关性最高的地方，但这是什么意思？是脸的中心吗？

如果有经验的人能够让我更加清楚，那就太好了。