操作系统：Centos7.4
参考：https://github.com/tensorflow/models/
1 环境安装
- 安装anaconda
[root@localhost home]# mkdir /home/anaconda_install
[root@localhost home]# cd /home/anaconda_install
[root@localhost anaconda_install]# wget https://repo.continuum.io/archive/Anaconda2-5.0.1-Linux-x86_64.sh
[root@localhost anaconda_install]# chmod 777 Anaconda2-5.0.1-Linux-x86_64.sh
[root@localhost anaconda_install]# ./Anaconda2-5.0.1-Linux-x86_64.sh
- conda安装pip：
[root@localhost anaconda_install]# conda install pip
- conda安装opencv3.0.0
[root@localhost anaconda_install]# conda install -c menpo opencv3
windows系统上对应命令： conda install -c https://conda.binstar.org/menpo opencv
- 安装tensorflow及依赖
[root@localhost anaconda_install]# mkdir -p /home/tensorflow/ #并把Object-Detector-App.tar.gz放到该目录解压）
[root@localhost anaconda_install]# cd /home/tensorflow/
[root@localhost tensorflow]# pip install tensorflow
[root@localhost tensorflow]# pip install pillow
[root@localhost tensorflow]# pip install pillow
[root@localhost tensorflow]# export PYTHONPATH=:/home/tensorflow/Object-Detector-App:/home/tensorflow/Object-Detector-App/slim
- 如果出现错误(importerror libpng12.so.0 cannot open shared object file no such file or directory)，安装png
yum install libpng12
- 如果出现`CXXABI_1.3.9' not found错误:
ImportError: /lib64/libstdc++.so.6: version `CXXABI_1.3.9' not found (required by /root/anaconda2/lib/python2.7/site-packages/matplotlib/_path.so)
解决方法：
[root@localhost Object-Detector-App]# find / -name "libstdc++.so.*"
find: ‘/run/user/1000/gvfs’: Permission denied
/root/anaconda2/pkgs/libstdcxx-ng-7.2.0-h7a57d05_2/lib/libstdc++.so.6.0.24
/root/anaconda2/pkgs/libstdcxx-ng-7.2.0-h7a57d05_2/lib/libstdc++.so.6
/root/anaconda2/pkgs/libstdcxx-ng-7.2.0-h7a57d05_2/x86_64-conda_cos6-linux-gnu/sysroot/lib/libstdc++.so.6
/root/anaconda2/pkgs/libstdcxx-ng-7.2.0-h7a57d05_2/x86_64-conda_cos6-linux-gnu/sysroot/lib/libstdc++.so.6.0.24
/root/anaconda2/lib/libstdc++.so.6
/root/anaconda2/lib/libstdc++.so.6.0.24
/root/anaconda2/x86_64-conda_cos6-linux-gnu/sysroot/lib/libstdc++.so.6
/root/anaconda2/x86_64-conda_cos6-linux-gnu/sysroot/lib/libstdc++.so.6.0.24
/usr/lib64/libstdc++.so.6
/usr/lib64/libstdc++.so.6.0.19
/usr/share/gdb/auto-load/usr/lib64/libstdc++.so.6.0.19-gdb.py
/usr/share/gdb/auto-load/usr/lib64/libstdc++.so.6.0.19-gdb.pyc
/usr/share/gdb/auto-load/usr/lib64/libstdc++.so.6.0.19-gdb.pyo
[root@localhost Object-Detector-App]# cp /root/anaconda2/pkgs/libstdcxx-ng-7.2.0-h7a57d05_2/lib/libstdc++.so.6.0.24 /lib64/
[root@localhost Object-Detector-App]# mv /lib64/libstdc++.so.6 /lib64/libstdc++.so.6.20171224
[root@localhost Object-Detector-App]# ln -s /lib64/libstdc++.so.6.0.24 /lib64/libstdc++.so.6
2 识别代码
2.1 识别的核心代码
import os
import cv2
import time
import argparse
import multiprocessing
import numpy as np
import tensorflow as tf
from utils.app_utils import FPS, WebcamVideoStream
from multiprocessing import Queue, Pool
from object_detection.utils import label_map_util
from object_detection.utils import visualization_utils as vis_util
CWD_PATH = os.getcwd()
# Path to frozen detection graph. This is the actual model that is used for the object detection.
MODEL_NAME = 'ssd_mobilenet_v1_coco_11_06_2017'
PATH_TO_CKPT = os.path.join(CWD_PATH, 'object_detection', MODEL_NAME, 'frozen_inference_graph.pb')
# List of the strings that is used to add correct label for each box.
PATH_TO_LABELS = os.path.join(CWD_PATH, 'object_detection', 'data', 'mscoco_label_map.pbtxt')
NUM_CLASSES = 90
# Loading label map
label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES,
use_display_name=True)
category_index = label_map_util.create_category_index(categories)
def detect_objects(image_np, sess, detection_graph):
# Expand dimensions since the model expects images to have shape: [1, None, None, 3]
image_np_expanded = np.expand_dims(image_np, axis=0)
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
# Each box represents a part of the image where a particular object was detected.
boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
# Each score represent how level of confidence for each of the objects.
# Score is shown on the result image, together with the class label.
scores = detection_graph.get_tensor_by_name('detection_scores:0')
classes = detection_graph.get_tensor_by_name('detection_classes:0')
num_detections = detection_graph.get_tensor_by_name('num_detections:0')
# Actual detection.
(boxes, scores, classes, num_detections) = sess.run(
[boxes, scores, classes, num_detections],
feed_dict={image_tensor: image_np_expanded})
# Visualization of the results of a detection.
vis_util.visualize_boxes_and_labels_on_image_array(
image_np,
np.squeeze(boxes),
np.squeeze(classes).astype(np.int32),
np.squeeze(scores),
category_index,
use_normalized_coordinates=True,
line_thickness=8)
return image_np
def worker(input_q, output_q):
# Load a (frozen) Tensorflow model into memory.
detection_graph = tf.Graph()
with detection_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def, name='')
sess = tf.Session(graph=detection_graph)
fps = FPS().start()
while True:
fps.update()
frame = input_q.get()
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
output_q.put(detect_objects(frame_rgb, sess, detection_graph))
fps.stop()
sess.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-src', '--source', dest='video_source', type=int,
default=0, help='Device index of the camera.')
parser.add_argument('-wd', '--width', dest='width', type=int,
default=480, help='Width of the frames in the video stream.')
parser.add_argument('-ht', '--height', dest='height', type=int,
default=360, help='Height of the frames in the video stream.')
parser.add_argument('-num-w', '--num-workers', dest='num_workers', type=int,
default=2, help='Number of workers.')
parser.add_argument('-q-size', '--queue-size', dest='queue_size', type=int,
default=5, help='Size of the queue.')
args = parser.parse_args()
logger = multiprocessing.log_to_stderr()
logger.setLevel(multiprocessing.SUBDEBUG)
input_q = Queue(maxsize=args.queue_size)
output_q = Queue(maxsize=args.queue_size)
pool = Pool(args.num_workers, worker, (input_q, output_q))
video_capture = WebcamVideoStream(src=args.video_source,
width=args.width,
height=args.height).start()
fps = FPS().start()
while True: # fps._numFrames < 120
frame = video_capture.read()
input_q.put(frame)
t = time.time()
output_rgb = cv2.cvtColor(output_q.get(), cv2.COLOR_RGB2BGR)
cv2.imshow('Video', output_rgb)
fps.update()
print('[INFO] elapsed time: {:.2f}'.format(time.time() - t))
if cv2.waitKey(1) & 0xFF == ord('q'):
break
fps.stop()
print('[INFO] elapsed time (total): {:.2f}'.format(fps.elapsed()))
print('[INFO] approx. FPS: {:.2f}'.format(fps.fps()))
pool.terminate()
video_capture.stop()
cv2.destroyAllWindows()
2.2 使用官方模型的运行结果
-
插入摄像头，查看设备号，我这边使用了4个摄像头，如下所示：
-
在/home/tensorflow/Object-Detector-App/目录下执行以下命令，打开一个摄像头(--source表示摄像头编号，我这里有0,1,2,3)：
python object_detection_app.py --source=0 --width=640 --height=480
-
效果如下(我这边同时打开了3个摄像头)：
3 使用自己的模型
目录结构：
3.1 准备工作
- Object Detection API
源码私聊可以私聊我（qq：479066524）。
源码包括：
sd_video_detect.py :视频检测
sd_pic_detect.py :图片识别
sd_train/ :训练脚本和资源
sd_model/ :训练后的模型
object_detection/ : 主要代码，可以从github拉取 - 训练自己的模型
如何训练自己的模型参考前面文章。自动标注图片程序可以私聊我。
我这里写了个脚本sd_train/train.sh:
#!/bin/bash
#set -x
#$1: model name
export PYTHONPATH=:/home/tensorflow/Object-Detector-App:/home/tensorflow/Object-Detector-App/slim
path=$(dirname `readlink -f $0`)
echo "path: $path"
echo "############step1: generate pascal_train.record############"
python $path/create_pascal_tf_record.py --data_dir=$path/$1/ \
--label_map_path=$path/$1/pascal_label_map.pbtxt \
--year=VOC2017 \
--set=train \
--output_path=$path/$1/pascal_train.record
if [ -f "$path/$1/pascal_train.record" ];then
echo "generate pascal_train.cord successfully: $path/$1/pascal_train.record"
else
echo "generate pascal_train.cord failed!!!"
exit -1
fi
sleep 1
echo "############step2: generate pascal_val.record############"
python $path/create_pascal_tf_record.py --data_dir=$path/$1/ \
--label_map_path=$path/$1/pascal_label_map.pbtxt \
--year=VOC2017 \
--set=val \
--output_path=$path/$1/pascal_val.record
if [ -f "$path/$1/pascal_val.record" ];then
echo "generate pascal_val.record successfully: $path/$1/pascal_val.record"
else
echo "generate pascal_val.record failed!!!"
exit -1
fi
sleep 1
echo "############step3: training############"
python $path/../object_detection/train.py --logtostderr \
--train_dir=$path/$1/output \
--pipeline_config_path=$path/$1/ssd_mobilenet_v1_pascal.config
- 生成可用的模型
我这边也写了个脚本sd_train/model.sh，供参考：
#!/bin/bash
set -x
export PYTHONPATH=:/home/tensorflow/Object-Detector-App:/home/tensorflow/Object-Detector-App/slim
#$1: model name
path=$(dirname `readlink -f $0`)
echo "path: $path"
#echo "############step1: generate model############"
#if [ -f "$path/$1/output/" ];then
# echo "generate pascal_train.cord successfully: $path/$1/pascal_train.record"
#else
# echo "generate pascal_train.cord failed!!!"
# exit -1
#fi
steps="`cat $path/$1/ssd_mobilenet_v1_pascal.config |grep num_steps |awk '{print $2}'`"
echo "train steps: $steps"
echo "############generate model############"
python $path/../object_detection/export_inference_graph.py --input_type image_tensor \
--pipeline_config_path $path/$1/ssd_mobilenet_v1_pascal.config \
--trained_checkpoint_prefix $path/$1/output/model.ckpt-$steps \
--output_directory $path/$1/savedModel
rm -rf $path/../sd_model/$1
cp -rf $path/$1/savedModel $path/../sd_model/$1
cp -rf $path/$1/ssd_mobilenet_v1_pascal.config $path/../sd_model/$1/
cp -rf $path/$1/pascal_label_map.pbtxt $path/../sd_model/$1/
echo "############completed! path: $path/../sd_model/$1/"
调用脚本后，生成的模型在sd_model/目录下：
- 使用模型
我的检测程序sd_video_detect.py：
import os
import cv2
import time
import argparse
import multiprocessing
import numpy as np
import tensorflow as tf
from utils.app_utils import FPS, WebcamVideoStream
from multiprocessing import Queue, Pool
from object_detection.utils import label_map_util
from object_detection.utils import visualization_utils as vis_util
CWD_PATH = os.getcwd()
parser = argparse.ArgumentParser()
parser.add_argument('-cn', '--classnum', dest='classnum', type=int,
default=90, help='Classes num of the model.')
parser.add_argument('-model', '--model', dest='model', type=str,
default='ssd_mobilenet_v1_coco_11_06_2017', help='the model name you want to run.')
parser.add_argument('-src', '--source', dest='video_source', type=int,
default=0, help='Device index of the camera.')
parser.add_argument('-wd', '--width', dest='width', type=int,
default=480, help='Width of the frames in the video stream.')
parser.add_argument('-ht', '--height', dest='height', type=int,
default=360, help='Height of the frames in the video stream.')
parser.add_argument('-num-w', '--num-workers', dest='num_workers', type=int,
default=2, help='Number of workers.')
parser.add_argument('-q-size', '--queue-size', dest='queue_size', type=int,
default=5, help='Size of the queue.')
args = parser.parse_args()
# Path to frozen detection graph. This is the actual model that is used for the object detection.
MODEL_NAME = args.model
PATH_TO_CKPT = os.path.join(CWD_PATH, 'sd_model', MODEL_NAME, 'frozen_inference_graph.pb')
# List of the strings that is used to add correct label for each box.
PATH_TO_LABELS = os.path.join(CWD_PATH, 'sd_model', MODEL_NAME, 'pascal_label_map.pbtxt')
NUM_CLASSES = args.classnum
# Loading label map
label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES,
use_display_name=True)
category_index = label_map_util.create_category_index(categories)
def detect_objects(image_np, sess, detection_graph):
# Expand dimensions since the model expects images to have shape: [1, None, None, 3]
image_np_expanded = np.expand_dims(image_np, axis=0)
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
# Each box represents a part of the image where a particular object was detected.
boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
# Each score represent how level of confidence for each of the objects.
# Score is shown on the result image, together with the class label.
scores = detection_graph.get_tensor_by_name('detection_scores:0')
classes = detection_graph.get_tensor_by_name('detection_classes:0')
num_detections = detection_graph.get_tensor_by_name('num_detections:0')
# Actual detection.
(boxes, scores, classes, num_detections) = sess.run(
[boxes, scores, classes, num_detections],
feed_dict={image_tensor: image_np_expanded})
# Visualization of the results of a detection.
vis_util.visualize_boxes_and_labels_on_image_array(
image_np,
np.squeeze(boxes),
np.squeeze(classes).astype(np.int32),
np.squeeze(scores),
category_index,
use_normalized_coordinates=True,
line_thickness=8)
return image_np
def worker(input_q, output_q):
# Load a (frozen) Tensorflow model into memory.
detection_graph = tf.Graph()
with detection_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def, name='')
sess = tf.Session(graph=detection_graph)
fps = FPS().start()
while True:
fps.update()
frame = input_q.get()
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
output_q.put(detect_objects(frame_rgb, sess, detection_graph))
fps.stop()
sess.close()
if __name__ == '__main__':
#parser = argparse.ArgumentParser()
#parser.add_argument('-cn', '--classnum', dest='classnum', type=int,
# default=90, help='Classes num of the model.')
#parser.add_argument('-model', '--model', dest='model', type=string,
# default='ssd_mobilenet_v1_coco_11_06_2017', help='the model name you want to run.')
#parser.add_argument('-src', '--source', dest='video_source', type=int,
# default=0, help='Device index of the camera.')
#parser.add_argument('-wd', '--width', dest='width', type=int,
# default=480, help='Width of the frames in the video stream.')
#parser.add_argument('-ht', '--height', dest='height', type=int,
# default=360, help='Height of the frames in the video stream.')
#parser.add_argument('-num-w', '--num-workers', dest='num_workers', type=int,
# default=2, help='Number of workers.')
#parser.add_argument('-q-size', '--queue-size', dest='queue_size', type=int,
# default=5, help='Size of the queue.')
#args = parser.parse_args()
logger = multiprocessing.log_to_stderr()
logger.setLevel(multiprocessing.SUBDEBUG)
input_q = Queue(maxsize=args.queue_size)
output_q = Queue(maxsize=args.queue_size)
pool = Pool(args.num_workers, worker, (input_q, output_q))
video_capture = WebcamVideoStream(src=args.video_source,
width=args.width,
height=args.height).start()
fps = FPS().start()
while True: # fps._numFrames < 120
frame = video_capture.read()
input_q.put(frame)
t = time.time()
output_rgb = cv2.cvtColor(output_q.get(), cv2.COLOR_RGB2BGR)
cv2.imshow('Video', output_rgb)
fps.update()
print('[INFO] elapsed time: {:.2f}'.format(time.time() - t))
if cv2.waitKey(1) & 0xFF == ord('q'):
break
fps.stop()
print('[INFO] elapsed time (total): {:.2f}'.format(fps.elapsed()))
print('[INFO] approx. FPS: {:.2f}'.format(fps.fps()))
pool.terminate()
video_capture.stop()
cv2.destroyAllWindows()
调用命令：
python sd_video_detect.py --classnum=2 \
--model=ssd_model_2017_12_15 \
--source=0 \
--width=640 \
--height=480
classnum: 表示模型中，物品种类数，我这边是两个；
model: 表示使用什么模型；
source: 表示使用的视频源；
width: 像素宽;
height: 像素高;
结果如下图：
红茶1
雪梨1
雪梨2
需要源码的可以Q我：479066524
网友评论