numpy: detect consecutive 1 in an array

Question

I want to detect consecutive spans of 1's in a numpy array. Indeed, I want to first identify whether the element in an array is in a span of a least three 1's. For example, we have the following array a:

    import numpy as np
    a = np.array([1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0])

Then the following 1's in bold are the elements satisfy the requirement.

[1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0]

Next, if two spans of 1's are separated by at most two 0's, then the two spans make up a longer span. So the above array is charaterized as

[1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0]

In other words, for the original array as input, I want the output as follows:

    [True, True, True, True, True, True, True, False, False, False, False, False, True, True, True, True, True, True, True, True, True, True, False]

I have been thinking of an algorithm to implement this function, but all the one I come up with seems to complicated. So I would love to know better ways to implement this -- it would be greatly appreciated if someone can help me out.

Update:

I apologize that I did not make my question clear. I want to identify 3 or more consecutive 1's in the array as a span of 1's, and any two spans of 1's with only one or two 0's in between are identified, along with the separating 0's, as a single long span. My goal can be understood in the following way: if there are only one or two 0's between spans of 1's, I consider those 0's as errors and are supposed to be corrected as 1's.

@ritesht93 provided an answer that almost gives what I want. However, the current answer does not identify the case when there are three spans of 1's that are separated by 0's, which should be identified as one single span. For example, for the array

    a2 = np.array([0, 1, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 0])

we should receive the output

    [False,  True,  True,  True,  True,  True,  True,  True,  True,
   True,  True,  True,  True,  True, False, False,  False, False,
   False,  True,  True,  True,  True,  True, False]

Update 2:

I was greatly inspired by and found the algorithm based on regular expression is easiest to implement and to understand -- though I am not sure about the efficient compared to other methods. Eventually I used the following method.

    lst = np.array([0, 1, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 0])
    lst1 = re.sub(r'1{3,}', lambda x:'c'*len(x.group()), ''.join(map(str, lst)))
    print lst1

which identified spans of 1's

    0ccc0ccc00cccc00100ccccc0

and then connect spans of 1's

    lst2 = re.sub(r'c{1}0{1,2}c{1}', lambda x:'c'*len(x.group()), ''.join(map(str, lst1)))
    print lst2

which gives

    0ccccccccccccc00100ccccc0

The final result is given by

    np.array(list(lst2)) == 'c'

    array([False,  True,  True,  True,  True,  True,  True,  True,  True,
    True,  True,  True,  True,  True, False, False, False, False,
   False,  True,  True,  True,  True,  True, False])

Answer 1

We could solve it with a combination of binary dilation and erosion to get past the first stage and then binary closing to get the final output, like so -

from scipy.ndimage.morphology import binary_erosion,binary_dilation,binary_closing

K = np.ones(3,dtype=int) # Kernel
b = binary_dilation(binary_erosion(a,K),K)
out = binary_closing(b,K) | b

Sample runs

Case #1 :

In [454]: a
Out[454]: array([1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0])

In [456]: out
Out[456]: 
array([ True,  True,  True,  True,  True,  True,  True, False, False,
       False, False, False,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True, False], dtype=bool)

Case #2 :

In [460]: a
Out[460]: 
array([0, 1, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 0])

In [461]: out
Out[461]: 
array([False,  True,  True,  True,  True,  True,  True,  True,  True,
        True,  True,  True,  True,  True, False, False, False, False,
       False,  True,  True,  True,  True,  True, False], dtype=bool)