Waterpixel 非常高效的超像素分割方法

【实例简介】

Waterpixel 非常高效的超像素分割方法

【实例截图】

【核心代码】

def demo_m_waterpixels(imin,  step,  d_weight, filter_ori):
    """
    Compute m-waterpixels, i.e. superpixels based on the watershed transformation.
    Flooding starts form the best minimum of each cell of a regular grid.
    The gradient used to be flooded is regularized using the distance function to these minima.
    Cells of the grid are chosen here to be squares.
    
    Input :
    imin (image UINT8): original image, to be segmented into superpixels
    step (UINT8) : grid step, i.e. distance between two cell centers of the grid
    d_weight (UINT8) : constant to be multiplied with function distance before addition to gradient image.
        If d_weight <=0, then only the gradient is taken into account.
    filter_ori (BOOLEAN) : do we filter the input image?

    Output:
    image (UINT16) : labelled superpixels
    image (UINT8) : distance weighted gradient function
    image (UINT16) : minima used in the computation
    """
    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------   
    # Connexity:
    basicse = sp.CrossSE()
    gridse = sp.SquSE()

    # Ori filtering
    if filter_ori is True:
        my_area_filtering(imin, step*step/16)
        imin.show()
    # Gradient computation
    im_grad = my_gradient(imin, basicse,  True)
    im_grad.show()   
    ## Pool of working images:
    imwrk0 = sp.Image(im_grad)
    imwrk1 = sp.Image(im_grad,  "UINT16")
    imwrk2 = sp.Image(im_grad,  "UINT16")

    # Compute cell centers and cells
    size = imin.getSize()
    im_markers, im_cells = im_labelled_square_grid_points(size, step, step/6)

    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------
    ## Choice of the markers : one per grid cell
    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------

    # Step 1 : Computation of the minima of the gradient
    im_minima = sp.Image(im_grad)
    sp.minima(im_grad, im_minima, basicse)
    #Step 2 : Imposing minimum distance between minima (= Removing minima candidates which fall on cell margins )
    sp.test(im_cells, im_minima, 0, imwrk0)
    sp.label(imwrk0, imwrk1, basicse)
    #Step 3 : Evaluation of the importance of minima ( = computation of their surfacic extinction)
    im_minima_val = sp.Image(imwrk1)
    sp.watershedExtinction( im_grad, imwrk1, im_minima_val, "a", basicse)
    # Step 4 : Taking back at value 2 the minima of null-volumic extinction
    sp.test( imwrk0,  2,  0,  imwrk2)
    sp.test( im_minima_val,  im_minima_val, imwrk2, im_minima_val )
    # Step 5 : Coping with the potential absence of minimum in cells (= choosing the minimum value inside this cell as its marker if necessary)
    imwrk00=sp.Image(imwrk0)
    blobs = sp.computeBlobs(im_cells)
    sp.test(im_cells, im_grad, 0, imwrk00)
    minVals = sp.measMinVals(imwrk00, blobs)
    sp.applyLookup(im_cells, minVals, imwrk0)
    sp.test(imwrk00==imwrk0, 1, 0, imwrk1)
    maxVals = sp.measMaxVals(im_minima_val, blobs)
    sp.applyLookup(im_cells, maxVals, imwrk2)
    sp.test(imwrk2, im_minima_val, imwrk1, im_minima_val)
    # Step 6 : Selection of one marker per cell
    one_min_per_grid_cell(im_cells, blobs, im_minima_val, basicse)

    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------
    ## Spatial regularization of the gradient
    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------  
    
    #Step 1 : Computation of the distance function to the markers
    immask = sp.Image(im_markers, "UINT8")
    sp.test(im_minima_val, 0, 1, immask)
    imdist = sp.Image(immask, "UINT8")
    sp.dist(immask, imdist, gridse)
    #Step 2 : Adding the distance function to the gradient
    if d_weight > 0:
        weight = d_weight * float(2)/step
        sp.mul(imdist, weight, imdist)
        sp.add(imdist, im_grad, im_grad)

    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------
    ## Superpixel segmentation : watershed transformation, flooding from selected minima on the regularized gradient
    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------
    sp.copy(im_minima_val, imwrk1)
    sp.label(imwrk1,  im_minima_val, basicse)
    imout = sp.Image(im_minima_val)
    sp.basins(im_grad, im_minima_val, imout, basicse)
    ##-----------------------------------------------------------------------------------------
    ##-----------------------------------------------------------------------------------------  
    return imout,  im_grad,  im_minima_val