Numpy sum messing up adding negative numbers-CodePudding

I'm trying to code a simple convolution loop for image processing. It seems to work fine when the kernel adds up to 1 i.e smoothing filter.

But when using an edge detection filter, weird values appear:

  for i in range(img.shape[1]):
    index = []
    for j in range(order):
      index.append(i j)
    array_aux=img_processed[:,index]
    img_final[:,i]=np.sum(array_aux*kernel_x,axis=1)

I'm applying a 1D kernel that im extending to the full height of the picture and convolving by multiplying that extended kernel with the columns needed and adding up all values. I´m using cv2.BORDER_REFLECT_101 padding:

  Image:                Kernel:    Final Image:

  [ 85  85  85  85  85] [ 1  0 -1] [  0   0   0   0   0]
  [ 85  85  85  85  85] [ 1  0 -1] [  0   0   0   0   0]
  [ 84  84  84  84  84] [ 1  0 -1] [  0   0   0   0   0]
           ...              ...             ... 
  [106 136 179 170 152] [ 1  0 -1] [  0 183 222  27   6]
  [113 129 172 175 153] [ 1  0 -1] [  0 197 210  19  30]
  [123 125 168 183 156] [ 1  0 -1] [  0 211 198  12  57]

as you can see while the first three row seem to be fine, the last 3 are wrong as negative number are expected. What's going on?

CodePudding user response：

Actually, your issue seems to be img_final, which I suspect is also of dtype uint8.

If you do an operation between a float and an integer, the result will correctly be a float. But when you try to put a float in a uint8 array, it will get cast to that.

>>> img_final = np.zeros((3, 3), dtype=np.uint8)
>>> img_final[0,0] = -2.5
>>> img_final
array([[254,   0,   0],
       [  0,   0,   0],
       [  0,   0,   0]], dtype=uint8)

Just set the correct type for img_final and it should be fine. If you want to handle positive and negative integers, int32 should suffice.