def main():
parser = setup_arg_parse()
args = parser.parse_args()
imglsh = ImageLSH(args.L, args.k)
reduced_data, image_ids = imglsh.load_data()
idx_structure, points_dict = imglsh.create_index_structure()
for key, val in idx_structure.items():
idx_structure[key] = list(val)
lsh_points = {}
for key, val in points_dict.items():
lsh_points[str(key)] = val
idx_structure['_id'] = "index"
lsh_points['_id'] = "points"
mongo_client = connect_to_db()
for img, data in zip(image_ids, reduced_data):
mongo_client.mwdb_project.image_features.update_one(
{'imageName': img}, {"$set": {
"HOG_reduced": data.tolist()
}})
mongo_client.mwdb_project.images_index.insert_one(lsh_points)
w_length = {'_id': "w_length", "w_length": imglsh.w_length}
mongo_client.mwdb_project.images_index.insert_one(w_length)
config = {'k': args.k, 'L': args.L, '_id': "config"}
mongo_client.mwdb_project.images_index.insert_one(config)
with open('index.json', 'w') as fp:
json.dump(idx_structure, fp)
mongo_client.close()
def classify_with_probablistic_feedback(query_vector, similar_images,
relevant_images):
mongo_client = connect_to_db()
value = 0
N = len(similar_images) * len(similar_images[0])
R = len(relevant_images) * len(relevant_images[0])
similarity_dict = convert_to_dict(similar_images)
reordered_images = []
for img in relevant_images:
di = similarity_dict[img]
image_vector = list(
mongo_client.mwdb_project.image_features.find(
{'imageName': img}))[0]["HOG_reduced"]
ni = count_occurances(image_vector, query_vector)
ri = count_relevant_occurances(relevant_images, image_vector)
pi = (ri + (ni / N)) / (R + 1)
ui = (ni - ri + (ni / N)) / (N - R + 1)
value -= di * np.log(pi * (1 - ui) / ui * (1 - pi))
reordered_images.append((img, value))
similarity_dict.pop(img)
mongo_client.close()
for k, v in similarity_dict.items():
reordered_images.append((k, v))
return sorted(reordered_images, key=lambda tup: tup[1])
def count_relevant_occurances(relevant_images, image_vector):
mongo_client = connect_to_db()
count = 0
for img in relevant_images:
vector = list(
mongo_client.mwdb_project.image_features.find(
{'imageName': img}))[0]["HOG_reduced"]
count += count_occurances(vector, image_vector)
mongo_client.close()
return count
def get_vectors_for_images(images):
mongo_client = connect_to_db()
result = []
for image in images:
result.append(
np.array(
list(
mongo_client.mwdb_project.image_features.find(
{'imageName': image}))[0]["HOG_reduced"]))
mongo_client.close()
return np.array(result)
def main():
parser = setup_arg_parse()
args = parser.parse_args()
mongo_client = connect_to_db()
config = list(mongo_client.mwdb_project.images_index.find({'_id' : "config"}))[0]
with open('index.json', 'r') as fp:
idx_structure = json.load(fp)
points_dict = list(mongo_client.mwdb_project.images_index.find({'_id': "points"}, {"_id" : 0}))[0]
w_length = list(mongo_client.mwdb_project.images_index.find({"_id" : "w_length"}))[0]["w_length"]
lsh_points = {}
for k, v in points_dict.items():
lsh_points[int(k)] = v
print(config)
query_image_details = list(mongo_client.mwdb_project.image_features.find({"imageName": args.query_image_id}))[0]
query_vector = query_image_details["HOG_reduced"]
imglsh = ImageLSH(config["k"], config["L"])
imglsh.load_index_structure(idx_structure, lsh_points, w_length)
similarity_scores, total_images_considered, unique_images_considered = imglsh.find_similar_images(query_vector, args.t, mongo_client)
similar_images = []
for score in similarity_scores:
image = {}
image["distance_score"] = score[1]
image["imageName"] = score[0]
image["image_path"] = list(mongo_client.mwdb_project.image_features.find({"imageName" : score[0]}))[0]["image_path"]
similar_images.append(image)
mongo_client.close()
print("Number of unique images considered = {0}".format(total_images_considered))
print("Number of overall Images considerd = {0}".format(unique_images_considered))
plot_results(similar_images, query_image_details["image_path"])
def get_seed_matrix(label):
mongo_client = connect_to_db()
images = mongo_client.mwdb_project.image_features.find({}, {
"imageName": 1,
"_id": 0,
"aspectOfHand": 1
})
seed_list = []
count = 0
for img in images:
if "aspectOfHand" in img:
if label in img["aspectOfHand"].lower():
count += 1
seed_list.append(1)
else:
seed_list.append(0)
else:
seed_list.append(0)
seed_matrix = np.array(seed_list)
seed_matrix = seed_matrix / np.sum(seed_matrix)
return seed_matrix
def main():
parser = setup_arg_parse()
args = parser.parse_args()
populate_database(args)
model = "CM"
dorsal_data_matrix, _ = get_data_matrix(
model, convert_label_to_filterstring("dorsal"))
palmar_data_matrix, _ = get_data_matrix(
model, convert_label_to_filterstring("palmar"))
dorsal_labels = np.zeros((dorsal_data_matrix.shape[0], 1))
palmar_labels = np.ones((palmar_data_matrix.shape[0], 1))
labels = np.append(dorsal_labels, palmar_labels, axis=0)
combined_data = np.append(dorsal_data_matrix, palmar_data_matrix, axis=0)
#reduced_data = reduce_dimensions_svd(combined_data, 20)
reduced_data, v_matrix = reduce_dimensions_svd(combined_data,
20,
get_v=True)
dx, ddx, labels, d_labels = train_test_split(reduced_data,
labels,
test_size=0.1,
random_state=42)
reduced_data = np.append(dx, ddx, axis=0)
labels = np.append(labels, d_labels, axis=0)
labeled_data = np.append(reduced_data, labels, axis=1)
testing_images, test_image_ids = enumerate_files_in_dir(args.test_folder)
test_dataset = []
for test_image, image_id in zip(testing_images, test_image_ids):
#test_dataset.append(np.array(extract_hog_features(test_image)))
test_dataset.append(np.array(extract_color_moments(test_image)))
test_dataset = np.array(test_dataset)
#reduced_test_dataset = reduce_dimensions_svd(test_dataset, 20)
reduced_test_dataset = np.matmul(test_dataset, v_matrix)
mongo_client = connect_to_db()
actual_labels = get_actual_labels_from_csv(args.labels_csv, test_image_ids)
predicted = []
if args.classifier == "DT":
model = DecisionTreeClassifier()
model.fit(labeled_data)
results = model.transform(reduced_test_dataset)
for test_image_id, result in zip(test_image_ids, results):
if result == 0:
label = "dorsal"
elif result == 1:
label = "palmar"
predicted.append((test_image_id, label))
print("{0} - {1}".format(test_image_id, label))
elif args.classifier == "SVM":
clf = SupportVectorMachine(kernel=rbf_kernel, power=4, coef=1)
training_labels = labels[:]
# SVM needs labels to be 1, and -1
training_labels[training_labels == 0] = -1
clf.fit(reduced_data, training_labels)
values = clf.predict(reduced_test_dataset)
print(values)
for test_image_id, result in zip(test_image_ids, values):
if result == 1:
label = "palmar"
else:
label = "dorsal"
predicted.append((test_image_id, label))
print("{0} - {1}".format(test_image_id, label))
elif args.classifier == "PPR":
args.k = 15
function_val = "manhattan"
#process_all_images(args.train_folder, "CM")
#process_all_images(args.test_folder, "CM")
outgoing_img_graph, image_ids = create_similarity_graph(
args.k, function_val, "CM")
transition_matrix = get_transition_matrix(outgoing_img_graph, args.k)
seed_matrix_dorsal = get_seed_matrix("dorsal")
seed_matrix_palmar = get_seed_matrix("palmar")
dorsal_pagerank = compute_pagerank(transition_matrix,
seed_matrix_dorsal)
palmar_pagerank = compute_pagerank(transition_matrix,
seed_matrix_palmar)
dorsal_pagerank_dict = {
x: y
for x, y in zip(image_ids, dorsal_pagerank)
}
palmar_pagerank_dict = {
x: y
for x, y in zip(image_ids, palmar_pagerank)
}
predicted = label_images(dorsal_pagerank_dict, palmar_pagerank_dict,
test_image_ids)
print(get_accuracy(actual_labels, predicted))
mongo_client.close()
def main():
parser = setup_arg_parse()
args = parser.parse_args()
mongo_client = connect_to_db()
config = list(
mongo_client.mwdb_project.images_index.find({'_id': "config"}))[0]
with open('index.json', 'r') as fp:
idx_structure = json.load(fp)
points_dict = list(
mongo_client.mwdb_project.images_index.find({'_id': "points"},
{"_id": 0}))[0]
w_length = list(
mongo_client.mwdb_project.images_index.find({"_id": "w_length"
}))[0]["w_length"]
lsh_points = {}
for k, v in points_dict.items():
lsh_points[int(k)] = v
print(config)
query_image_details = list(
mongo_client.mwdb_project.image_features.find(
{"imageName": args.query_image_id.split("/")[-1]})
)[0] #a change was made here cuz my database does not contain imageName as a whole path
query_vector = query_image_details["HOG_reduced"]
imglsh = ImageLSH(config["k"], config["L"])
imglsh.load_index_structure(idx_structure, lsh_points, w_length)
#full_data_matrix, image_ids = get_data_matrix("HOG")
#full_data_matrix = reduce_dimensions_lda(full_data_matrix, 256)
# print(full_data_matrix.shape)
similarity_scores, total_images_considered, unique_images_considered = imglsh.find_similar_images(
query_vector, args.t, mongo_client)
jmp_matrix = None
while True:
pprint(similarity_scores)
similar_images = []
for score in similarity_scores:
image = {}
image["imageName"] = score[0]
image["image_path"] = list(
mongo_client.mwdb_project.image_features.find(
{"imageName": score[0]}))[0]["image_path"]
similar_images.append(image)
plot_results(
similar_images, query_image_details["image_path"]
) #work has to be done here, we need R and IR to be created by this
R, IR = get_feedback_ids()
if args.clf == "SVM":
similarity_scores = classify_with_svm(query_vector,
similarity_scores, R, IR)
elif args.clf == "PPR":
similarity_scores, jmp_matrix = classify_with_ppr(
query_vector, similarity_scores, R, IR)
elif args.clf == "DT":
similarity_scores = classify_with_dt(query_vector,
similarity_scores, R, IR)
else:
similarity_scores = classify_with_probablistic_feedback(
query_vector, similarity_scores, R)