def filter_candidate_pool(self, candidate_pool, char):
"""Filter candidate pool based on secondary distance metrics and
selected threshold. Secondary distance is calculated between the image
of the anchor character and the image of the candidate character.
Args:
candidate_pool: Set of Char, the set of possible confusables.
char: Char, single character whose corresponding image must exists
in self.img_dir.
"""
# Get secondary distance metrics
image_metrics = Distance(self.img_format).get_metrics()
if self.secondary_distance_type not in image_metrics.keys():
raise ValueError(
"Expect secondary_distance_type to be one of {}.".format(
image_metrics.keys()))
secondary_dis = image_metrics[self.secondary_distance_type]
# Filter candidate pool to get confusables
confusables = []
for candidate in candidate_pool:
if ord(char) == ord(candidate):
continue
dis = calculate_from_path(secondary_dis, self._label_img_map[char],
self._label_img_map[candidate])
if dis <= self.secondary_filter_threshold:
confusables.append((candidate, dis))
return confusables
def get_candidate_pool_for_char(self, char):
"""Obtain the candidates for confusables for specified 'char'. Use
the reduced representations generated by PCA and select candidates based
on primary distance metrics.
Args:
char: Char, single character, must exists in self.labels.
Returns:
candidate_pool: Set of Char, the set of possible confusables.
candidate_dis: Dict, mapping from candidates to their respective
distances.
"""
# Get character index in labels and embeddings
idx = self.labels.index(char)
# Get a pool of possible candidates for secondary filter
candidate_pool = set()
# Store distances between all confusables and anchor
candidate_dis = dict()
for embs in self._reps:
# Get embedding anchor to compare with others
emb_anchor = embs[idx]
# Get primary distance metrics
embedding_metrics = Distance(ImgFormat.EMBEDDINGS).get_metrics()
if self.primary_distance_type not in embedding_metrics.keys():
raise ValueError(
"Expect primary_distance_type to be one of {}.".format(
embedding_metrics.keys()))
primary_dis = embedding_metrics[self.primary_distance_type]
# Get distance from anchor embedding to all other embeddings
distances = []
for emb in embs:
distances.append(primary_dis(emb_anchor, emb))
label_dis_pairs = list(zip(self.labels, distances))
# Get top n candidates using the primary distance metric
top_n = []
for label, dis in label_dis_pairs:
if len(top_n) < self.n_candidates:
# Append reversed tuple for sorting
bisect.insort(top_n, (dis, label))
else:
if dis < top_n[self.n_candidates - 1][0]:
# If the distance is lower than the largest of the
# candidates we only keep top N
bisect.insort(top_n, (dis, label))
top_n = top_n[:self.n_candidates - 1]
# Store all candidate distances
candidate_dis["PCA" + str(embs.shape[1])] = top_n
candidate_pool = candidate_pool.union(
set([entry[1] for entry in top_n]))
return candidate_pool, candidate_dis