def search_t_binary(haze_patch, comparator, start_t, end_t, A, step_tol):
'''
The main difference from method2 is the step size is halved after
each step and the remaining part is changed accordingly
'''
# assuming start_t < end_t
t_begin = start_t
t_end = end_t
conf = 1
# abs() is not required as t_end > t_begin. Always.
while (t_end - t_begin) > step_tol and \
(t_end > t_begin):
t_mid = (t_begin + t_end) / 2.0
p_d = dehaze_patch(haze_patch, t_mid, A)
[a, b] = comp_patches(comparator, haze_patch, p_d)
if a > b:
# the dehazed one is bad
# no need to search below t_mid
t_begin = t_mid
else:
t_end = t_mid
# need to handle out of range values and t_end < t_begin case
if t_begin > t_end:
conf = 0
# this does not have check, as it is done in search_t_binary2
# whether the t tries to go out of range
t_mid = (t_begin + t_end) / 2.0
return (t_mid, conf)
# the t_comp's should not be too close
dist = pdist(np.reshape(t_comp, (-1, 1)),
metric='minkowski',
p=1)
if np.any(dist < t_step):
continue
in_index = idx * n_in_bad_p[idx]
bad_t[in_index:in_index + n_in_bad_p[idx]] = t_comp
break
# print bad_t
assert (np.all(bad_t < t_g))
for t_good in good_t:
d_p = dehaze_patch(haze_patch, t_good, A)
add_patch(d_p, haze_patch, X_all, Y_all, X_idx)
A_all[X_idx, :, :, :] = A
X_idx += 1
assert (np.sum((haze_patch - d_p)**2) != 0)
for t_bad in bad_t:
d_p = dehaze_patch(haze_patch, t_bad, A)
add_patch(haze_patch, d_p, X_all, Y_all, X_idx)
A_all[X_idx, :, :, :] = A
X_idx += 1
assert (np.sum((haze_patch - d_p)**2) != 0)
t_all[i, j, 0] = t_g