def find_tall_glyphs(self, stdev=20):
from gamera import stats
tall = []
for i in range(len(self.glyphs)):
g = self.glyphs[i]
if stats.samplestdev([g.nrows, self.avg_glyph_height]) > stdev:
tall.append(i)
return tall
def find_tall_glyphs(self, stdev=20):
from gamera import stats
tall = []
for i in range(len(self.glyphs)):
g = self.glyphs[i]
if stats.samplestdev([g.nrows, self.avg_glyph_height]) > stdev:
tall.append(i)
return tall
def make_subtrees_stddev(graph, ratio, distance, relabel=1, lab="cluster."):
cur_label = 0
remove = []
for edge in graph.get_edges():
lengths = []
path = {}
#print node().get_main_id(), edge.cost
get_lengths(edge.from_node, distance, lengths, 0, path)
lengths.remove(edge.cost)
#print lengths
if not (len(lengths) > 1):
continue
mean = stats.mean(lengths)
stdev2 = stats.samplestdev([mean, edge.cost])
#print mean, stdev2, edge.cost, len(lengths)
if stdev2 > ratio:
#graph.remove_edge(edge)
remove.append(edge)
for edge in remove:
graph.remove_edge(edge)
if relabel:
cur_label = 0
for node in graph.get_nodes():
node().classify_manual("")
for node in graph.get_nodes():
if node().get_main_id() == "":
label(graph, node, lab, cur_label)
cur_label += 1
nodes = []
for node in graph.get_nodes():
nodes.append(node())
return nodes
def make_subtrees_stddev(graph, ratio, distance, relabel=1, lab="cluster."):
cur_label = 0
remove = []
for edge in graph.get_edges():
lengths = []
path = {}
#print node().get_main_id(), edge.cost
get_lengths(edge.from_node, distance, lengths, 0, path)
lengths.remove(edge.cost)
#print lengths
if not (len(lengths) > 1):
continue
mean = stats.mean(lengths)
stdev2 = stats.samplestdev([mean, edge.cost])
#print mean, stdev2, edge.cost, len(lengths)
if stdev2 > ratio:
#graph.remove_edge(edge)
remove.append(edge)
for edge in remove:
graph.remove_edge(edge)
if relabel:
cur_label = 0
for node in graph.get_nodes():
node().classify_manual("")
for node in graph.get_nodes():
if node().get_main_id() == "":
label(graph, node, lab, cur_label)
cur_label += 1
nodes = []
for node in graph.get_nodes():
nodes.append(node())
return nodes
def test():
glyphs = gamera_xml.glyphs_from_xml(r"C:\Documents and Settings\Karl MacMillan\Desktop\test\prod.xml")
glyphs = strip_small_categories(glyphs)
from gamera.plugins import features
k = knn.kNN()
print k.features
features.generate_features_list(glyphs, k.feature_functions)
print "Getting gstats"
graph_stats = get_graph_stats(glyphs, k)
gstats = knn.get_glyphs_stats(glyphs)
max_dist = max_distance(glyphs, k)
print max_dist
file = open("results.txt", "w")
global_max = [[],[]]
local_max = [[],[]]
all = [[],[]]
graph = [[],[]]
gr_ccorrect = 0
gr_icorrect = 0
for x in glyphs:
local_max_dist = local_max_distance(glyphs, x, k)
ans = k.classify_with_images(glyphs, x, 1)
file.write(ans[0][1] + ",")# + str(ans[0][0]) + ",")
correct = 0
if x.get_main_id() == ans[0][1]:
file.write("1,")
correct = 1
else:
file.write("0,")
g = 1.0 - (ans[0][0] / max_dist)
global_max[correct].append(g)
file.write(str(g) + ",")
l = 1.0 - (ans[0][0] / local_max_dist)
local_max[correct].append(l)
file.write(str(l) + ",")
a = stats.samplestdev([ans[0][0],gstats[ans[0][1]][1]])
all[correct].append(a)
file.write(str(a) + ",")
gr = stats.samplestdev([ans[0][0],graph_stats[ans[0][1]]])
if (gr <= 1 and correct):
gr_ccorrect += 1
if (gr > 1 and not correct):
gr_icorrect += 1
graph[correct].append(gr)
file.write(str(gr))
file.write("\n")
print "num correct: %d num incorrect: %d" % (len(global_max[1]), len(global_max[0]))
print "confidence %f %f %f" % (((gr_ccorrect + gr_icorrect) / float(len(glyphs))),
gr_ccorrect / float(len(glyphs) - len(global_max[0])),
gr_icorrect / float(len(glyphs) - len(global_max[1])))
cgm = -1
igm = -1
cgs = -1
igs = -1
if (len(global_max[0])):
igm = stats.mean(global_max[0])
igs = stats.samplestdev(global_max[0])
if (len(global_max[1])):
cgm = stats.mean(global_max[1])
cgs = stats.samplestdev(global_max[1])
clm = -1
ilm = -1
cls = -1
ils = -1
if (len(local_max[0])):
ilm = stats.mean(local_max[0])
ils = stats.samplestdev(local_max[0])
if (len(local_max[1])):
clm = stats.mean(local_max[1])
cls = stats.samplestdev(local_max[1])
cam = -1
iam = -1
cas = -1
ias = -1
if (len(all[0])):
iam = stats.mean(all[0])
ias = stats.samplestdev(all[0])
if (len(all[1])):
cam = stats.mean(all[1])
cas = stats.samplestdev(all[1])
cgraphm = -1
igraphm = -1
cgraphs = -1
igraphs = -1
if (len(graph[0])):
igraphm = stats.mean(graph[0])
igraphs = stats.samplestdev(graph[0])
if (len(graph[1])):
cgraphm = stats.mean(graph[1])
cgraphs = stats.samplestdev(graph[1])
print "global correct avg: %f stdev: %f incorrect avg: %f stddev: %f" % (cgm, cgs, igm, igs)
print "local correct avg: %f stdev: %f incorrect avg: %f stddev: %f" % (clm, cls, ilm, ils)
print "all correct avg: %f stdev: %f incorrect avg: %f stddev: %f" % (cam, cas, iam, ias)
print "graph correct avg: %f stdev: %f incorrect avg: %f stddev: %f" % (cgraphm, cgraphs, igraphm, igraphs)
def otsu_threshold(p):
l = len(p)
mu_T = 0.0
for i in range(l):
mu_T += i * p[i]
sigma_T = 0.0
for i in range(l):
sigma_T += (i-mu_T)*(i-mu_T)*p[i]
k_low = 0
while (p[k_low] == 0) and (k_low < (l - 1)):
k_low += 1
k_high = l - 1
while (p[k_high] == 0) and (k_high > 0):
k_low += 1
k_high -= 1
criterion = 0.0
thresh = 127
omega_k = 0.0
mu_k = 0.0
k = k_low
while k <= k_high:
omega_k += p[k]
mu_k += k*p[k]
expr_1 = (mu_T*omega_k - mu_k)
sigma_b_k = expr_1 * expr_1 / (omega_k*(1-omega_k))
if (criterion < sigma_b_k/sigma_T):
criterion = sigma_b_k/sigma_T
thresh = k;
k += 1
return thresh
graph_l = graph[0][:]
graph_l.extend(graph[1])
graph_l.sort()
threshold = stats.mean(graph_l)
print "threshold: " + str(threshold)
num_wrong = 0
for x in graph[0]:
if x < threshold:
num_wrong += 1
print num_wrong, num_wrong / float(len(graph[0])) * 100
num_wrong = 0
for x in graph[1]:
if x >= threshold:
num_wrong += 1
print num_wrong, num_wrong / float(len(graph[1])) * 100
graph_l = all[0][:]
graph_l.extend(all[1])
graph_l.sort()
threshold = stats.mean(graph_l)
print "threshold: " + str(threshold)
num_wrong = 0
for x in graph[0]:
if x < threshold:
num_wrong += 1
print num_wrong, num_wrong / float(len(graph[0])) * 100
num_wrong = 0
for x in graph[1]:
if x >= threshold:
num_wrong += 1
print num_wrong, num_wrong / float(len(graph[1])) * 100
graph_l = local_max[0][:]
graph_l.extend(local_max[1])
graph_l.sort()
threshold = stats.mean(graph_l)
print "threshold: " + str(threshold)
num_wrong = 0
for x in graph[0]:
if x < threshold:
num_wrong += 1
print num_wrong, num_wrong / float(len(graph[0])) * 100
num_wrong = 0
for x in graph[1]:
if x >= threshold:
num_wrong += 1
print num_wrong, num_wrong / float(len(graph[1])) * 100
def test():
glyphs = gamera_xml.glyphs_from_xml(
r"C:\Documents and Settings\Karl MacMillan\Desktop\test\prod.xml")
glyphs = strip_small_categories(glyphs)
from gamera.plugins import features
k = knn.kNN()
print(k.features)
features.generate_features_list(glyphs, k.feature_functions)
print("Getting gstats")
graph_stats = get_graph_stats(glyphs, k)
gstats = knn.get_glyphs_stats(glyphs)
max_dist = max_distance(glyphs, k)
print(max_dist)
file = open("results.txt", "w")
global_max = [[], []]
local_max = [[], []]
all = [[], []]
graph = [[], []]
gr_ccorrect = 0
gr_icorrect = 0
for x in glyphs:
local_max_dist = local_max_distance(glyphs, x, k)
ans = k.classify_with_images(glyphs, x, 1)
file.write(ans[0][1] + ",") # + str(ans[0][0]) + ",")
correct = 0
if x.get_main_id() == ans[0][1]:
file.write("1,")
correct = 1
else:
file.write("0,")
g = 1.0 - (ans[0][0] / max_dist)
global_max[correct].append(g)
file.write(str(g) + ",")
l = 1.0 - (ans[0][0] / local_max_dist)
local_max[correct].append(l)
file.write(str(l) + ",")
a = stats.samplestdev([ans[0][0], gstats[ans[0][1]][1]])
all[correct].append(a)
file.write(str(a) + ",")
gr = stats.samplestdev([ans[0][0], graph_stats[ans[0][1]]])
if (gr <= 1 and correct):
gr_ccorrect += 1
if (gr > 1 and not correct):
gr_icorrect += 1
graph[correct].append(gr)
file.write(str(gr))
file.write("\n")
print("num correct: %d num incorrect: %d" %
(len(global_max[1]), len(global_max[0])))
print("confidence %f %f %f" %
(((gr_ccorrect + gr_icorrect) / float(len(glyphs))),
gr_ccorrect / float(len(glyphs) - len(global_max[0])),
gr_icorrect / float(len(glyphs) - len(global_max[1]))))
cgm = -1
igm = -1
cgs = -1
igs = -1
if (len(global_max[0])):
igm = stats.mean(global_max[0])
igs = stats.samplestdev(global_max[0])
if (len(global_max[1])):
cgm = stats.mean(global_max[1])
cgs = stats.samplestdev(global_max[1])
clm = -1
ilm = -1
cls = -1
ils = -1
if (len(local_max[0])):
ilm = stats.mean(local_max[0])
ils = stats.samplestdev(local_max[0])
if (len(local_max[1])):
clm = stats.mean(local_max[1])
cls = stats.samplestdev(local_max[1])
cam = -1
iam = -1
cas = -1
ias = -1
if (len(all[0])):
iam = stats.mean(all[0])
ias = stats.samplestdev(all[0])
if (len(all[1])):
cam = stats.mean(all[1])
cas = stats.samplestdev(all[1])
cgraphm = -1
igraphm = -1
cgraphs = -1
igraphs = -1
if (len(graph[0])):
igraphm = stats.mean(graph[0])
igraphs = stats.samplestdev(graph[0])
if (len(graph[1])):
cgraphm = stats.mean(graph[1])
cgraphs = stats.samplestdev(graph[1])
print("global correct avg: %f stdev: %f incorrect avg: %f stddev: %f" %
(cgm, cgs, igm, igs))
print("local correct avg: %f stdev: %f incorrect avg: %f stddev: %f" %
(clm, cls, ilm, ils))
print("all correct avg: %f stdev: %f incorrect avg: %f stddev: %f" %
(cam, cas, iam, ias))
print("graph correct avg: %f stdev: %f incorrect avg: %f stddev: %f" %
(cgraphm, cgraphs, igraphm, igraphs))
def otsu_threshold(p):
l = len(p)
mu_T = 0.0
for i in range(l):
mu_T += i * p[i]
sigma_T = 0.0
for i in range(l):
sigma_T += (i - mu_T) * (i - mu_T) * p[i]
k_low = 0
while (p[k_low] == 0) and (k_low < (l - 1)):
k_low += 1
k_high = l - 1
while (p[k_high] == 0) and (k_high > 0):
k_low += 1
k_high -= 1
criterion = 0.0
thresh = 127
omega_k = 0.0
mu_k = 0.0
k = k_low
while k <= k_high:
omega_k += p[k]
mu_k += k * p[k]
expr_1 = (mu_T * omega_k - mu_k)
sigma_b_k = expr_1 * expr_1 / (omega_k * (1 - omega_k))
if (criterion < sigma_b_k / sigma_T):
criterion = sigma_b_k / sigma_T
thresh = k
k += 1
return thresh
graph_l = graph[0][:]
graph_l.extend(graph[1])
graph_l.sort()
threshold = stats.mean(graph_l)
print("threshold: " + str(threshold))
num_wrong = 0
for x in graph[0]:
if x < threshold:
num_wrong += 1
print(num_wrong, num_wrong / float(len(graph[0])) * 100)
num_wrong = 0
for x in graph[1]:
if x >= threshold:
num_wrong += 1
print(num_wrong, num_wrong / float(len(graph[1])) * 100)
graph_l = all[0][:]
graph_l.extend(all[1])
graph_l.sort()
threshold = stats.mean(graph_l)
print("threshold: " + str(threshold))
num_wrong = 0
for x in graph[0]:
if x < threshold:
num_wrong += 1
print(num_wrong, num_wrong / float(len(graph[0])) * 100)
num_wrong = 0
for x in graph[1]:
if x >= threshold:
num_wrong += 1
print(num_wrong, num_wrong / float(len(graph[1])) * 100)
graph_l = local_max[0][:]
graph_l.extend(local_max[1])
graph_l.sort()
threshold = stats.mean(graph_l)
print("threshold: " + str(threshold))
num_wrong = 0
for x in graph[0]:
if x < threshold:
num_wrong += 1
print(num_wrong, num_wrong / float(len(graph[0])) * 100)
num_wrong = 0
for x in graph[1]:
if x >= threshold:
num_wrong += 1
print(num_wrong, num_wrong / float(len(graph[1])) * 100)