def dynamic_modification_timer(mutex):
time.sleep(3600)
mutex.acquire()
flag = False
global gallery_feature
global image_paths
'''
增加
直接读取并加入到特征集中；
并将图片加密写入数据库
'''
base_path = './static/image_database/'
# Prepare model. 加载预训练的model
# #可删除部分
# model = load_model(pretrained_model='./retrieval/models/net_best.pth', use_gpu=True)
# print("Model load successfully!")
# gallery_feature = np.load('./retrieval/models/gallery_feature.npy') #ndarray
# image_paths = np.load('./retrieval/models/image_paths.npy')
# gallery_feature = torch.from_numpy(gallery_feature)
# image_paths = image_paths.tolist()
# #到这
update_path = './static/update_images/'
for path in os.listdir(update_path):
data_loader = load_data(data_path=update_path+path+'/',
batch_size=1,
shuffle=False,
transform='default',
)
increase_feature, increase_image_paths = extract_feature(model=model,
dataloaders=data_loader) # torch.Size([59, 2048]) # tensor list
# 加密并加入数据库(修改成AES)
if len(increase_image_paths) != 0:
flag = True
print('添加成功')
conn = dbinfo()
conncur = conn.cursor()
try:
for i in range(0, len(increase_image_paths)):
path = increase_image_paths[i]
entend_path = path[path.rfind('/') + 1:]
fp = open(path, 'rb')
AES_data = aes.encrypt(fp.read())
#base64_data = mybase64.encode(fp.read())
fp.close()
try:
sql_insertimage = "replace into image_AES (file_name,image_value) VALUE (%s, %s) "
conncur.execute(sql_insertimage, (entend_path, AES_data))
seatdic = conncur.fetchall()
conn.commit()
conn.close()
except pymysql.Error as e:
print("Error %d %s" % (e.args[0], e.args[1]))
sys.exit(1)
new_path = base_path + entend_path
shutil.copyfile(path, new_path)
os.remove(path)
increase_image_paths[i] = new_path
except IOError as e:
print("Error %d %s" % (e.args[0], e.args[1]))
sys.exit(1)
gallery_feature = np.load('./retrieval/models/'+path+'_feature.npy')
gallery_feature = torch.from_numpy(gallery_feature)
image_paths = np.load('./retrieval/models/'+path+'_paths.npy')
image_paths = image_paths.tolist()
gallery_feature = torch.cat((gallery_feature, increase_feature), 0)
image_paths.extend(increase_image_paths)
# print(gallery_feature)
# print(image_paths)
# np.save('./retrieval/models/gallery_feature.npy', gallery_feature.numpy())
# np.save('./retrieval/models/image_paths.npy', np.array(image_paths))
'''
删除
在数据库中删除，然后用它的文件名在image_path中查找得到下标，删除向量即可；
'''
# Prepare model. 加载预训练的model
# #可删除部分
# model = load_model(pretrained_model='./retrieval/models/net_best.pth', use_gpu=True)
# print("Model load successfully!")
# gallery_feature = np.load('./retrieval/models/gallery_feature.npy') #ndarray
# image_paths = np.load('./retrieval/models/image_paths.npy')
# gallery_feature = torch.from_numpy(gallery_feature)
# image_paths = image_paths.tolist()
# #到这
delete_path = './static/delete_images/'
for path in os.listdir(delete_path):
data_loader = load_data(data_path=delete_path + path + '/',
batch_size=1,
shuffle=False,
transform='default',
)
delete_feature, delete_image_paths = extract_feature(model=model,
dataloaders=data_loader) # torch.Size([59, 2048]) # tensor list
if len(delete_image_paths) != 0:
print('删除成功')
conn = dbinfo()
conncur = conn.cursor()
del_sql = 'delete from image_AES where file_name = %s'
flag = True
for i in range(0, len(delete_image_paths)):
path = delete_image_paths[i]
os.remove(path)
delete_image_paths[i] = path[path.rfind('/') + 1:]
# 删除数据库数据
try:
conncur.executemany(del_sql, delete_image_paths)
conn.commit()
except:
conn.rollback()
conn.close()
del_list = []
for i in range(0, len(image_paths)):
path = image_paths[i]
extend_path = path[path.rfind('/') + 1:]
if extend_path in delete_image_paths:
del_list.append(i)
os.remove(path)
gallery_feature = np.load('./retrieval/models/' + path + '_feature.npy')
gallery_feature = torch.from_numpy(gallery_feature)
image_paths = np.load('./retrieval/models/' + path + '_paths.npy')
image_paths = image_paths.tolist()
gallery_feature = gallery_feature.numpy()
gallery_feature = np.delete(gallery_feature, del_list, axis=0)
image_paths = np.delete(image_paths, del_list)
gallery_feature = torch.from_numpy(gallery_feature)
image_paths = image_paths.tolist()
if flag == True:
np.save('./retrieval/models/gallery_feature.npy', gallery_feature.numpy())
np.save('./retrieval/models/image_paths.npy', np.array(image_paths))
else:
print('未修改')
mutex.release()
global thread
# 重复构造定时器
thread = threading.Thread(target=dynamic_modification_timer, args=(mutex, ))
thread.start()
model = load_model(pretrained_model=os.path.join('./DealNet/checkpoint', 'DealNet', 'ckpt.t7'), use_gpu=True)
print("Model load successfully!")
local_dir = './static/image_database/'
# Extract database features.
# 在数据库图片不改变的情况下 选择是否保存特征向量 以节约时间
if args.update:
# Extract database features.
for dir_name in os.listdir(local_dir):
# Prepare data set.
#print(dir_name)
data_loader = load_data(data_path=local_dir + dir_name+'/',
batch_size=1,
shuffle=False,
transform='default',
)
gallery_feature, image_paths = extract_feature(model=model, dataloaders=data_loader) # torch.Size([59, 2048])
#print(gallery_feature)
np.save('./retrieval/models/'+dir_name+'_feature.npy', gallery_feature.numpy())
np.save('./retrieval/models/'+dir_name+'_paths.npy', np.array(image_paths))
# else:
# gallery_feature = np.load('./retrieval/models/gallery_feature.npy')
# gallery_feature = torch.from_numpy(gallery_feature)
# image_paths = np.load('./retrieval/models/image_paths.npy')
# image_paths = image_paths.tolist()
print("extract_feature save successfully!")
# mybase64 = MyBase64()
# s = "vwxrstuopq34567ABCDEFGHIJyz012PQRSTKLMNOZabcdUVWXYefghijklmn89+/"
# mybase64.__init__(s)