TensorFlow2学习二十、预训练模型FasterRCNN+InceptionResNetV2目标检测
一、概念
目标检测是在真实场景中寻找类似车辆、人类等物体的过程，一般可以在给定的图像中寻找多个目标。它可以用在图像检索、安防、自动驾驶（ADAS）等系统。
目标可以有以下几种方式：
- 基于特征的目标检测
- Viola Jones目标检测
- 基于HOG特征的SVM分类
- 深度学习
一些目标检测应用场景
-
DeepFace :Facebook用来人脸识别
-
Google facial recognition system
-
人数统计
-
工业产品质量检查
-
自动驾驶
-
安防 人脸解锁
二、目标检测流程
不同目标检测算法的流程有所不同，但原理都差不多。下图是一个示例流程：
特征提取是关键动作，可以通过MatLab,OpenCV,Viola Jones 或深度学习来实现特征的提取。
三、实现
本文使用google训练好的模型 FasterRCNN+InceptionResNetV2或ssd+mobilenet V2来进行目标检测测试。
本文环境
- TensorFlow2.0
- 谷歌Colab
1. 导入包
#@title Imports and function definitions
# Currently %tensorflow_version 2.x installs beta1, which doesn't work here.
# %tensorflow_version can likely be used after 2.0rc0
!pip install tf-nightly-gpu-2.0-preview
# For running inference on the TF-Hub module.
import tensorflow as tf
import tensorflow_hub as hub
# For downloading the image.
import matplotlib.pyplot as plt
import tempfile
from six.moves.urllib.request import urlopen
from six import BytesIO
# For drawing onto the image.
import numpy as np
from PIL import Image
from PIL import ImageColor
from PIL import ImageDraw
from PIL import ImageFont
from PIL import ImageOps
# For measuring the inference time.
import time
# Check available GPU devices.
print("The following GPU devices are available: %s" % tf.test.gpu_device_name())
2.
# 显示图片
def display_image(image):
fig = plt.figure(figsize=(20, 15))
plt.grid(False)
plt.imshow(image)
# 下载、预处理图片
def download_and_resize_image(url, new_width=256, new_height=256,
display=False):
# 下载图片
_, filename = tempfile.mkstemp(suffix=".jpg")
response = urlopen(url)
image_data = response.read()
image_data = BytesIO(image_data)
pil_image = Image.open(image_data)
# 统一大小
pil_image = ImageOps.fit(pil_image, (new_width, new_height), Image.ANTIALIAS)
pil_image_rgb = pil_image.convert("RGB")
pil_image_rgb.save(filename, format="JPEG", quality=90)
print("Image downloaded to %s." % filename)
if display:
display_image(pil_image)
return filename
# 在图像上画框
def draw_bounding_box_on_image(image,
ymin,
xmin,
ymax,
xmax,
color,
font,
thickness=4,
display_str_list=()):
"""Adds a bounding box to an image."""
draw = ImageDraw.Draw(image)
im_width, im_height = image.size
(left, right, top, bottom) = (xmin * im_width, xmax * im_width,
ymin * im_height, ymax * im_height)
draw.line([(left, top), (left, bottom), (right, bottom), (right, top),
(left, top)],
width=thickness,
fill=color)
# If the total height of the display strings added to the top of the bounding
# box exceeds the top of the image, stack the strings below the bounding box
# instead of above.
display_str_heights = [font.getsize(ds)[1] for ds in display_str_list]
# Each display_str has a top and bottom margin of 0.05x.
total_display_str_height = (1 + 2 * 0.05) * sum(display_str_heights)
if top > total_display_str_height:
text_bottom = top
else:
text_bottom = bottom + total_display_str_height
# Reverse list and print from bottom to top.
for display_str in display_str_list[::-1]:
text_width, text_height = font.getsize(display_str)
margin = np.ceil(0.05 * text_height)
draw.rectangle([(left, text_bottom - text_height - 2 * margin),
(left + text_width, text_bottom)],
fill=color)
draw.text((left + margin, text_bottom - text_height - margin),
display_str,
fill="black",
font=font)
text_bottom -= text_height - 2 * margin
# 图像上加注释
def draw_boxes(image, boxes, class_names, scores, max_boxes=10, min_score=0.1):
"""Overlay labeled boxes on an image with formatted scores and label names."""
colors = list(ImageColor.colormap.values())
try:
font = ImageFont.truetype("/usr/share/fonts/truetype/liberation/LiberationSansNarrow-Regular.ttf",
25)
except IOError:
print("Font not found, using default font.")
font = ImageFont.load_default()
for i in range(min(boxes.shape[0], max_boxes)):
if scores[i] >= min_score:
ymin, xmin, ymax, xmax = tuple(boxes[i])
display_str = "{}: {}%".format(class_names[i].decode("ascii"),
int(100 * scores[i]))
color = colors[hash(class_names[i]) % len(colors)]
image_pil = Image.fromarray(np.uint8(image)).convert("RGB")
draw_bounding_box_on_image(
image_pil,
ymin,
xmin,
ymax,
xmax,
color,
font,
display_str_list=[display_str])
np.copyto(image, np.array(image_pil))
return image
3. 下载 OpenImagesV4图片
image_url = "https://farm1.staticflickr.com/4032/4653948754_c0d768086b_o.jpg" #@param
downloaded_image_path = download_and_resize_image(image_url, 1280, 856, True)
4. 选择一种模型：
- FasterRCNN+InceptionResNet V2：准确度高
- ssd+mobilenet V2:小而且快
module_handle = "https://tfhub.dev/google/faster_rcnn/openimages_v4/inception_resnet_v2/1"
#@param ["https://tfhub.dev/google/openimages_v4/ssd/mobilenet_v2/1", "https://tfhub.dev/google/faster_rcnn/openimages_v4/inception_resnet_v2/1"]
detector = hub.load(module_handle).signatures['default']
5. 定义函数加载图片
def load_img(path):
img = tf.io.read_file(path)
img = tf.image.decode_jpeg(img, channels=3)
return img
6. 检测函数
def run_detector(detector, path):
img = load_img(path)
converted_img = tf.image.convert_image_dtype(img, tf.float32)[tf.newaxis, ...]
start_time = time.time()
result = detector(converted_img)
end_time = time.time()
result = {key:value.numpy() for key,value in result.items()}
print("Found %d objects." % len(result["detection_scores"]))
print("Inference time: ", end_time-start_time)
image_with_boxes = draw_boxes(
img.numpy(), result["detection_boxes"],
result["detection_class_entities"], result["detection_scores"])
display_image(image_with_boxes)
7. 执行一个检测
run_detector(detector, downloaded_image_path)
8. 测试更多图像
image_urls = ["https://farm7.staticflickr.com/8092/8592917784_4759d3088b_o.jpg",
"https://farm6.staticflickr.com/2598/4138342721_06f6e177f3_o.jpg",
"https://c4.staticflickr.com/9/8322/8053836633_6dc507f090_o.jpg"]
for image_url in image_urls:
start_time = time.time()
image_path = download_and_resize_image(image_url, 640, 480)
run_detector(detector, image_path)
end_time = time.time()
print("Inference time:")
