def __init__(self):
self.ims = ImsES(Elasticsearch())
self.CVAlgorithm = CVModule()
# Init by configure_runtime.py
self.db_path = annoy_index_db_path
self.image_search = ImageSearch(self.db_path)
self.imageTrain = ImageTrain(self.db_path)
def __init__(self, db_name):
"""初始化
Args:
db_name (string): 需要加载的数据库名称
"""
self.MIN_MATCH_COUNT = 10
self.shape = (100, 128)
self.cvmodule = CVModule()
self.annoyindx = AnnoyIndex_driver(db_name)
self.ims = ImsES(Elasticsearch())
try:
self.annoyindx.loadDb()
except BaseException as e:
pass
class Image_Del_API(Resource):
def __init__(self):
self.ims = ImsES(Elasticsearch())
self.CVAlgorithm = CVModule()
# Init by configure_runtime.py
self.db_path = annoy_index_db_path
self.image_search = ImageSearch(self.db_path)
self.imageTrain = ImageTrain(self.db_path)
def delete(self):
img, self.args = get_image(self.CVAlgorithm)
img = self.CVAlgorithm.crop_center(img, dim=dim_800x800)
kps, des = self.CVAlgorithm.extract_feature(img)
# Init and load search algorithm
result_table = self.image_search.search_batch(des)
self.image_search.unload()
# Collection the id fields
result_ids_table = list()
[result_ids_table.append(result['id']) for result in result_table]
records = self.ims.search_multiple_record(result_ids_table)
# Annoy index data found, We also need to match the data one by one
for data_index in range(len(result_table)):
record = records[data_index]["_source"]
data = result_table[data_index]
RANSAC_percent = self.CVAlgorithm.findHomgraphy(
data['good'], kps, record['kps'])
# If it finds any very similar data, it will jump out of the test directly
if RANSAC_percent >= 0.5:
self.ims.delete_siginle_record({'id': record['id']})
# Loading already dataset and rebuild it!
already_dataset = self.ims.search_all_record()
for key in already_dataset:
self.imageTrain.addMarkerDes(key["_source"]["id"],
key["_source"]["des"])
# Rebuild index
if self.imageTrain.generateMarkerDB():
return {'msg': success_response}
else:
return {'msg': 'eror!'}
return {'msg': already_img_response}, 202
class ServerWorker(threading.Thread):
"""ServerWorker"""
def __init__(self, context):
threading.Thread.__init__(self)
self.context = context
self.CVModule = CVModule()
self.ims = ImsES(Elasticsearch())
def run(self):
worker = self.context.socket(zmq.XREQ)
worker.connect('inproc://backend')
print('Worker started')
recvCount = 0
while True:
data = worker.recv()
if len(data) > 5:
img = self.CVModule.read_base64(data)
crop_predict_img = self.CVModule.crop_center(
img, int(img.shape[0] * 0.8), int(img.shape[0] * 0.8))
kp, des = self.CVModule.extract_feature(crop_predict_img)
image_search = ImageSearch("cache/index.db")
result_table = image_search.search_batch(des)
if len(result_table) > 0:
record = self.ims.search_single_record(
{'id': result_table['id']})
if len(record) > 0:
record.pop('des')
result_table = self.merge_dicts(record, result_table)
worker.send(json.dumps(result_table).encode('utf8'))
else:
worker.send(json.dumps({}).encode('utf8'))
del data
worker.close()
def merge_dicts(self, dict1, dict2):
dict3 = dict1.copy()
dict3.update(dict2)
return dict3
class Image_Predict_API(Resource):
def __init__(self):
self.ims = ImsES(Elasticsearch())
self.CVAlgorithm = CVModule()
self.image_search = ImageSearch(annoy_index_db_path)
def post(self):
try:
img, self.args = get_image(self.CVAlgorithm)
img = self.CVAlgorithm.crop_center(img, dim=dim_800x800)
kps, des = self.CVAlgorithm.extract_feature(img)
result_table = self.image_search.search_batch(des)
# Collection the id fields
result_ids_table = list()
[result_ids_table.append(result['id']) for result in result_table]
records = self.ims.search_multiple_record(result_ids_table)
# Check the result length, when the result length is greater than 0, get the matching data
for data_index in range(len(result_table)):
record = records[data_index]["_source"]
data = result_table[data_index]
RANSAC_percent = self.CVAlgorithm.findHomgraphy(
data['good'], kps, record['kps'])
if len(record) > 0 and RANSAC_percent >= 0.5:
# Remove the field of des. Because the des field is storing the image description data
record.pop('des')
record.pop('kps')
data.pop('good')
data['confidence'] = RANSAC_percent
data = self.merge_dicts(data, record)
return data, 200
except Exception as BaseException:
print(BaseException)
pass
return {'data': '', 'message': 'Can not found!'}, 200
def merge_dicts(self, dict1, dict2):
dict3 = dict1.copy()
dict3.update(dict2)
return dict3
class ImageSearch():
def __init__(self, db_name):
"""初始化
Args:
db_name (string): 需要加载的数据库名称
"""
self.MIN_MATCH_COUNT = 10
self.shape = (100, 128)
self.cvmodule = CVModule()
self.annoyindx = AnnoyIndex_driver(db_name)
self.ims = ImsES(Elasticsearch())
try:
self.annoyindx.loadDb()
except BaseException as e:
pass
def find_vector(self, des):
"""通过向量匹配对应向量
Args:
des (vector): 需匹配向量
Returns:
vector: 匹配到的向量
"""
return self.annoyindx.find_vector(des)
def unload(self):
self.annoyindx.unload()
def get_item_vector_by_id(self, id):
"""通过数据库id获取该id对应的向量数据
Args:
id (int): 数据库id
Returns:
vector: 图像描述符向量
"""
return self.annoyindx.get_item_vector_by_id(id, self.shape)
def get_count(self):
"""获取当前annoy index数据库的个数
Returns:
int: 当前存在数据库内的数据个数
"""
return self.annoyindx.get_count()
def search_batch(self, targetVector):
"""批量检索相似向量
Args:
targetVector (向量): 图像描述符的向量
Returns:
Dict: 检索到最为匹配的图像数据
"""
kn_results = self.find_vector(targetVector)
result_table = list()
good = None
# terms检索 避免一次次检索浪费时间
data_caches_es = self.ims.search_multiple_record(kn_results)
try:
# Tips: 由于查询到的循序与Annoy Index查询到的顺序不一致，故使用ES返回数据id为配准
for data_index in range(len(kn_results)):
record = dict()
source = data_caches_es[data_index]['_source']
flatten_vector = source['des']
# 避免无法重塑形状
if len(flatten_vector) > 12800:
vector = self.annoyindx.reshape(
flatten_vector, (int(len(flatten_vector) / 128), 128))
else:
vector = self.annoyindx.reshape(flatten_vector, (100, 128))
good = self.cvmodule.match(targetVector, vector)
if len(good) >= self.MIN_MATCH_COUNT:
record['id'] = source['id']
record['matchscore'] = len(good)
record['good'] = good
result_table.append(record)
except BaseException as ex:
print(ex)
result_table.sort(key=self.result_sort, reverse=True)
return result_table
def result_sort(self, e):
return e['matchscore']
def __init__(self):
self.ims = ImsES(Elasticsearch())
self.CVAlgorithm = CVModule()
self.image_search = ImageSearch(annoy_index_db_path)
self.imageTrain = ImageTrain(annoy_index_db_path)
class Image_Train_API(Resource):
def __init__(self):
self.ims = ImsES(Elasticsearch())
self.CVAlgorithm = CVModule()
self.image_search = ImageSearch(annoy_index_db_path)
self.imageTrain = ImageTrain(annoy_index_db_path)
def post(self):
img, self.args = get_image(self.CVAlgorithm)
img = self.CVAlgorithm.crop_center(img, dim=dim_800x800)
if img.shape != dim_800x800:
return {'msg': 'Image size is not enough'}, 200
kps,des = self.CVAlgorithm.extract_feature(img)
# Init and load search algorithm
result_table = self.image_search.search_batch(des)
image_count = self.image_search.get_count()
self.image_search.unload()
if des is None:
return {'msg': 'Image description is not enough'}, 200
# Annoy index no data found
if len(result_table) == 0:
return self.build_record(image_count,kps,des)
# Annoy index data found, We also need to match the data one by one
for data in result_table:
record = self.ims.search_single_record({'id':data['id']})
RANSAC_percent = self.CVAlgorithm.findHomgraphy(data['good'], kps, record['kps'])
# If it finds any very similar data, it will jump out of the test directly
if RANSAC_percent > 0.5:
return {'data':'', 'msg': 'there are already image'}, 202
return self.build_record(image_count,kps,des)
def build_record(self,image_count,kps,des):
keypoint_serialize = [{'angle': k.angle, 'response': k.response,'octave':k.octave,'class_id':k.class_id,'pt':k.pt,'size':k.size} for k in kps]
metadata = self.args['metadata']
if metadata is None:
metadata = self.args['image'].filename
# Append data to already dataset
new_record = {'id': image_count,
'metadata':metadata,
'des': des.flatten(),
'kps': keypoint_serialize}
self.ims.insert_single_record(new_record, refresh_after=True)
# Loading already dataset
already_dataset = self.ims.search_all_record()
# Trainning it!
for key in already_dataset:
self.imageTrain.addMarkerDes(key["_source"]["id"],key["_source"]["des"])
# Rebuild index
if self.imageTrain.generateMarkerDB():
return {'msg': 'success'}
def __init__(self, context):
threading.Thread.__init__(self)
self.context = context
self.CVModule = CVModule()
self.ims = ImsES(Elasticsearch())