def random_prediction():
r = random.randint(0, TRAIN_PATH.shape[0] - 1)
img = np.asarray(
Image.open(
LIP_DATASET_SRC / TRAIN_PATH[r, 0]
)
)
print("Original:")
show_img(img)
rescaled_img = preprocess_image(img)
test_predict = model.predict(tf.expand_dims(rescaled_img, axis=0))
plot_parsing_map(test_predict, label2int=ATRDataset.LABEL2INT)
pred_mask = create_mask(test_predict)[0]
show_img(pred_mask)
import numpy as np
import matplotlib.pyplot as plt
import os
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
# 在推断模式中，将使用running_mean和running_variance作为bn层的mean和variance
tf.keras.backend.set_learning_phase(False)
# 加载模型
model = tf.saved_model.load('saved_model/overfit_model-1000e')
# 80个类别的名称
class_name = read_class_name('./model_data/coco_classes_chinese.txt')
# 图片名称
file = 'car2.png'
# 图片放缩成608,608,像素值归一化
image, image_data, image_shape = preprocess_image(img_path='train-set/origin/' + file)
# 进行目标检测算法
res_class, res_score, res_boxes = detect(image_data, model)
if len(res_boxes) > 0:
# image_shape 为原图大小高宽格式，
# 还原成相对于原图的位置
res_boxes = res_boxes * np.tile(list(image_shape), 2)
# 把框画在图片上
draw_boxes(image, res_score, res_boxes, res_class, class_name, generate_colors(class_name))
plt.imshow(image)
plt.show()
image.save('train-set/detected/'+file, quality=100)
else:
print('未检测到任何目标')
import numpy as np
import os
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
model = DarkNet()
# 编译网络，生成weights参数结构
model(tf.keras.layers.Input(shape=(608, 608, 3)))
# 解析权重参数
yolo_weights = parse_yolo_v2_model_weights(model.weight_shapes, 'yolov2_weights/yolov2.weights')
# 将权重参数导入模型
model.set_weights(yolo_weights)
# 80个类别的名称
class_name = read_class_name('./model_data/coco_classes_chinese.txt')
# 图片名称
f = 'car1.png'
# 图片放缩成608,608,像素值归一化
image, image_data, image_shape = preprocess_image(img_path='images/'+f)
# 进行目标检测算法
res_class, res_score, res_boxes = detect(image_data, model)
# image_shape 为原图大小高宽格式，
# 还原成相对于原图的位置
res_boxes = res_boxes * np.tile(list(image_shape), 2)
# 把框画在图片上
draw_boxes(image, res_score, res_boxes, res_class, class_name, generate_colors(class_name))
# plt.imshow(image)
# plt.show()
# 保存图片
image.save('images/out/' + f, quality=100)
model(tf.keras.layers.Input(shape=(608, 608, 3)))
print(model.summary())
# 解析权重参数
yolo_weights = parse_yolo_v2_model_weights(model.weight_shapes,
'yolov2_weights/yolov2.weights')
# 将权重参数导入模型
model.set_weights(yolo_weights)
# 80个类别的名称
class_name = read_class_name('./model_data/coco_classes_chinese.txt')
root_path = 'images'
# 检测文件夹下的所有图片，深度为1
for f in os.listdir(root_path):
nf = os.path.join(root_path, f)
if os.path.isfile(nf):
# 图片放缩成608,608,像素值归一化
image, image_data, image_shape = preprocess_image(img_path=nf)
# 进行目标检测算法
res_class, res_score, res_boxes = detect(image_data, model)
# image_shape 为原图大小高宽格式，
# 还原成相对于原图的位置
res_boxes = res_boxes * np.tile(list(image_shape), 2)
# 把框画在图片上
draw_boxes(image, res_score, res_boxes, res_class, class_name,
generate_colors(class_name))
# plt.imshow(image)
# plt.show()
# 保存图片
image.save('images/out/' + f, quality=100)
def predict(filename):
channel = grpc.insecure_channel("localhost:8500")
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
request = predict_pb2.PredictRequest()
# model_name
request.model_spec.name = "cropper_model"
# signature name, default is 'serving_default'
request.model_spec.signature_name = "serving_default"
start = time.time()
"""
=========================================
===== Crop and align id card image
=========================================
"""
filepath = app.config["IMAGE_UPLOADS"] + "/" + filename
# preprocess image
img, original_image, original_width, original_height = preprocess_image(
filepath, Cropper.TARGET_SIZE)
if img.ndim == 3:
img = np.expand_dims(img, axis=0)
# request to cropper model
request.inputs["input_1"].CopyFrom(
tf.make_tensor_proto(img, dtype=np.float32, shape=img.shape))
try:
result = stub.Predict(request, 10.0)
result = result.outputs["tf_op_layer_concat_14"].float_val
result = np.array(result).reshape((-1, 9))
except Exception as e:
print(e)
cropper = Cropper()
cropper.set_best_bboxes(result,
original_width=original_width,
original_height=original_height,
iou_threshold=0.5)
# respone to client if image is invalid
if not cropper.respone_client(threshold_idcard=0.8):
return render_template('upload_image_again.html')
cropper.set_image(original_image=original_image)
# output of cropper part
aligned_image = getattr(cropper, "image_output")
cv2.imwrite('app/static/aligned_images/' + filename, aligned_image)
aligned_image = cv2.cvtColor(aligned_image, cv2.COLOR_BGR2RGB)
"""
===========================================
==== Detect informations in aligned image
===========================================
"""
# preprocess aligned image
original_height, original_width, _ = aligned_image.shape
img = cv2.resize(aligned_image, Detector.TARGET_SIZE)
img = np.float32(img / 255.)
# model_name
request.model_spec.name = "detector_model"
# signature name, default is 'serving_default'
request.model_spec.signature_name = "serving_default"
if img.ndim == 3:
img = np.expand_dims(img, axis=0)
# new request to detector model
request.inputs["input_1"].CopyFrom(
tf.make_tensor_proto(img, dtype=np.float32, shape=img.shape))
try:
result = stub.Predict(request, 10.0)
result = result.outputs["tf_op_layer_concat_14"].float_val
result = np.array(result).reshape((-1, 13))
except Exception as e:
print(e)
detector = Detector()
detector.set_best_bboxes(result,
original_width=original_width,
original_height=original_height,
iou_threshold=0.5)
detector.set_info_images(original_image=aligned_image)
# output of detector part
info_images = getattr(detector, "info_images")
"""
=====================================
==== Reader infors from infors image
=====================================
"""
keys = list(info_images.keys())
keys.remove("thoi_han")
keys.remove("chan_dung")
infors = dict()
# init default value of quoc_tich, dan_toc
infors['quoc_tich'] = ""
infors['dan_toc'] = ""
if "quoc_tich" in keys:
infors['quoc_tich'] = ["Việt Nam"]
keys.remove("quoc_tich")
if "sex" in keys:
info_image = info_images["sex"]
infors["sex"] = list()
for i in range(len(info_image)):
img = info_image[i]['image']
s = reader.predict(img)
if "Na" in s:
infors["sex"].append("Nam")
else:
infors["sex"].append("Nữ")
keys.remove("sex")
if "dan_toc" in keys:
info_image = info_images["dan_toc"]
infors["dan_toc"] = list()
for i in range(len(info_image)):
img = info_image[i]['image']
s = reader.predict(img)
s = s.split(" ")[-1]
infors["dan_toc"].append(s)
keys.remove("dan_toc")
for key in keys:
infors[key] = list()
info_image = info_images[key]
for i in range(len(info_image)):
img = info_image[i]['image']
s = reader.predict(img)
infors[key].append(s)
que_quan_0 = infors['que_quan'][0]
que_quan_1 = ''
noi_thuong_tru_0 = infors['noi_thuong_tru'][0]
noi_thuong_tru_1 = ''
if len(infors['que_quan']) == 2:
que_quan_1 = infors['que_quan'][1]
if len(infors['noi_thuong_tru']) == 2:
noi_thuong_tru_1 = infors['noi_thuong_tru'][1]
print("total_time:{}".format(time.time() - start))
return render_template('predict.html',
id=infors['id'][0].replace(" ", ""),
full_name=infors['full_name'][0],
date_of_birth=infors['date_of_birth'][0],
sex=infors['sex'][0],
quoc_tich=infors['quoc_tich'],
dan_toc=infors['dan_toc'],
que_quan_0=que_quan_0,
que_quan_1=que_quan_1,
noi_thuong_tru_0=noi_thuong_tru_0,
noi_thuong_tru_1=noi_thuong_tru_1,
filename=str(filename))