【TensorFlow学习笔记】目标识别篇（五）：基于Faster-rcnn-tensorflow-python3.5-master打造属于自己的物体检测模型

2023-07-25 21:20:22

一、环境配置

win10系统，显卡GeForce GTX 960M；
TensorFlow-gpu 1.13.0-rc2，CUDA 10.0，Cudnn 7.4.2；
python 3.6

Tensorflow-gpu是在windows PowerShell里用pip安装的，同时安装一些必要的库：cython、easydict、matplotlib、python-opencv等，可直接使用pip安装或者下载相应的.whl离线文件安装。

Faster RCNN下载

下载地址https://github.com/dBeker/Faster-RCNN-TensorFlow-Python3.5

下载完成后，项目的根目录为：Faster-RCNN-TensorFlow-Python3.5-master

cd到Faster-RCNN-TensorFlow-Python3.5-master\data\coco\PythonAPI目录下，打开cmd，运行编译提供的代码：

python setup.py build_ext --inplace
python setup.py build_ext install

二、准备数据集

三、模型下载

VGG16模型的下载地址：http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz，也可去百度云盘下载，
链接：https://pan.baidu.com/s/11Ty10NJ-rgXkkvM92SVVKw ，提取码：d2jz
下载完后解压，文件重命名为vgg16.ckpt，如图2所示。新建文件夹imagenet_weights，把vgg16.ckpt放到imagenet_weights下，再将imagenet_weights文件夹复制到data文件夹下。文件夹目录：
Faster-RCNN-TensorFlow-Python3.5-master\data\imagenet_weights\vgg16.ckpt

resnet模型下载

四、训练模型

训练模型的参数可以在Faster-RCNN-TensorFlow-Python3.5-master\lib\config文件夹里的config.py修改，包括训练的总步数、权重衰减、学习率、batch_size等参数。

tf.app.flags.DEFINE_float('weight_decay', 0.0005, "Weight decay, for regularization")
tf.app.flags.DEFINE_float('learning_rate', 0.001, "Learning rate")
tf.app.flags.DEFINE_float('momentum', 0.9, "Momentum")
tf.app.flags.DEFINE_float('gamma', 0.1, "Factor for reducing the learning rate")
 
tf.app.flags.DEFINE_integer('batch_size', 128, "Network batch size during training")
tf.app.flags.DEFINE_integer('max_iters', 40000, "Max iteration")
tf.app.flags.DEFINE_integer('step_size', 30000, "Step size for reducing the learning rate, currently only support one step")
tf.app.flags.DEFINE_integer('display', 20, "Iteration intervals for showing the loss during training, on command line interface")
 
tf.app.flags.DEFINE_string('initializer', "truncated", "Network initialization parameters")
tf.app.flags.DEFINE_string('pretrained_model', "./data/imagenet_weights/vgg16.ckpt", "Pretrained network weights")
 
tf.app.flags.DEFINE_boolean('bias_decay', False, "Whether to have weight decay on bias as well")
tf.app.flags.DEFINE_boolean('double_bias', True, "Whether to double the learning rate for bias")
tf.app.flags.DEFINE_boolean('use_all_gt', True, "Whether to use all ground truth bounding boxes for training, "
                                                "For COCO, setting USE_ALL_GT to False will exclude boxes that are flagged as ''iscrowd''")
tf.app.flags.DEFINE_integer('max_size', 1000, "Max pixel size of the longest side of a scaled input image")
tf.app.flags.DEFINE_integer('test_max_size', 1000, "Max pixel size of the longest side of a scaled input image")
tf.app.flags.DEFINE_integer('ims_per_batch', 1, "Images to use per minibatch")
tf.app.flags.DEFINE_integer('snapshot_iterations', 5000, "Iteration to take snapshot")

参数调整完后，在Faster-RCNN-TensorFlow-Python3.5-master的目录下，运行 python train.py，就可以训练生成模型了。

模型训练结束后，在 Faster-RCNN-TensorFlow-Python3.5-master\default\voc_2007_trainval\default目录下可以看到训练的模型，一个迭代了40000次，迭代次数可在Faster-RCNN-TensorFlow-Python3.5-master\lib\config文件夹里的config.py修改。

在目录下新建output\vgg16\voc_2007_trainval\default文件，将训练生成的文件复制到该文件下，并改名如下:“vgg16.ckpt.meta”，如图4所示：

五、测试模型

对demo.py进行如下的修改

1、将NETS中的“vgg16_faster_rcnn_iter_70000.ckpt”改成“vgg16”，如下所示；

NETS = {‘vgg16’: (‘vgg16.ckpt’,), ‘res101’: (‘res101_faster_rcnn_iter_110000.ckpt’,)}
2、将DATASETS中的“voc_2007_trainval+voc_2012_trainval”改为“voc_2007_trainval”，如下所示；

DATASETS = {‘pascal_voc’: (‘voc_2007_trainval’,), ‘pascal_voc_0712’: (‘voc_2007_trainval’,)}
3、将def parse_args()函数的两个default分别改成vgg16和pascal_voc，如下所示；

def parse_args():
“”“Parse input arguments.”""
parser = argparse.ArgumentParser(description=‘Tensorflow Faster R-CNN demo’)
parser.add_argument(’–net’, dest=‘demo_net’, help=‘Network to use [vgg16 res101]’,
choices=NETS.keys(), default=‘vgg16’)
parser.add_argument(’–dataset’, dest=‘dataset’, help=‘Trained dataset [pascal_voc pascal_voc_0712]’,
choices=DATASETS.keys(), default=‘pascal_voc’)
args = parser.parse_args()

return args

填坑

1.运行train.py 报错No module named cython_bbox 这是由于fasterrcnn为基于pyhton3.5的代码，但是之前安装的python版本为3.6，解决方法很简单

Faster-RCNN：ModuleNotFoundError: No module named
‘lib.utils.cython_bbox’
【已解决】
https://blog.csdn.net/memories_sunset/article/details/82117628

2.error: Unable to find vcvarsall.bat 这个错误的原因为我的vs版本为2013，但是程序运行需要vs2015的环境，解决这个错误不需要重新安装vs2015.

只需要安装其building tools即可，下载链接在这里

https://devblogs.microsoft.com/python/unable-to-find-vcvarsall-bat/#comments

选择python3.5 or latter下载即可

下载之后直接安装，选择自定义，选择win10版本

安装过程中可能会提示安装包损坏，不用管，跳过即可

安装成功后，反过来去解决第一个问题

之后即可成功运行train.py来进行faster-rcnn训练了
https://blog.csdn.net/kellyroslyn/article/details/92799174

测试视频，代码如下，截取视频的一张图片如图6所示：

# -*- coding: utf-8 -*-
# --------------------------------------------------------
# Faster R-CNN
#author lk
# --------------------------------------------------------
"""
Demo script showing detections in videos.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
 
import argparse
import os
import cv2
import tensorflow as tf
from lib.config import config as cfg
from lib.utils.test import im_detect
from lib.utils.nms_wrapper import nms
from lib.utils.timer import Timer
from lib.nets.vgg16 import vgg16
import matplotlib.pyplot as plt
import numpy as np
import sys
import time
 
(major_ver, minor_ver, subminor_ver) = (cv2.__version__).split('.')
 
CLASSES = ('__background__',
           'aeroplane', 'bicycle', 'bird', 'boat',
           'bottle', 'bus', 'car', 'cat', 'chair',
           'cow', 'diningtable', 'dog', 'horse',
           'motorbike', 'person', 'pottedplant',
           'sheep', 'sofa', 'train', 'tvmonitor')
 
NETS = {'vgg16': ('vgg16.ckpt',), 'res101': ('res101_faster_rcnn_iter_110000.ckpt',)}
DATASETS = {'pascal_voc': ('voc_2007_trainval',), 'pascal_voc_0712': ('voc_2007_trainval',)}
 
def vis_detections(im, class_name, dets, thresh=0.5):
    """Draw detected bounding boxes."""
    inds = np.where(dets[:, -1] >= thresh)[0]
    if len(inds) == 0:
        return
 
    im = im[:, :, (2, 1, 0)]
    fig, ax = plt.subplots(figsize=(12, 12))
    ax.imshow(im, aspect='equal')
    for i in inds:
        bbox = dets[i, :4]
        score = dets[i, -1]
 
        ax.add_patch(
            plt.Rectangle((bbox[0], bbox[1]),
                          bbox[2] - bbox[0],
                          bbox[3] - bbox[1], fill=False,
                          edgecolor='red', linewidth=3.5)
        )
        ax.text(bbox[0], bbox[1] - 2,
                '{:s} {:.3f}'.format(class_name, score),
                bbox=dict(facecolor='blue', alpha=0.5),
                fontsize=14, color='white')
 
    ax.set_title(('{} detections with '
                  'p({} | box) >= {:.1f}').format(class_name, class_name,
                                                  thresh),
                 fontsize=14)
 
    plt.axis('off')
    plt.tight_layout()
    plt.draw()
 
def vis_detections_video(im, class_name, dets, thresh=0.5):
    """Draw detected bounding boxes."""
    #np.where判断语句
    inds = np.where(dets[:, -1] >= thresh)[0]
 
    if len(inds) == 0:
        return im
 
    for i in inds:
        bbox = dets[i, :4]
        score = dets[i, -1]
        cv2.rectangle(im, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (0, 0, 255), 2)
        cv2.rectangle(im, (int(bbox[0]), int(bbox[1] - 20)), (int(bbox[0] + 200), int(bbox[1])), (10, 10, 10), -1)
 
        cv2.putText(im, '{:s} {:.3f}'.format(class_name, score), (int(bbox[0]), int(bbox[1] - 2)),
                    cv2.FONT_HERSHEY_SIMPLEX, .75, (0, 0, 255))
 
    return im
 
def demo(net, im):
    """Detect object classes in an image using pre-computed object proposals."""
    global frameRate
    global fps
 
    # Detect all object classes and regress object bounds
 
    timer = Timer()
    timer.tic()
    scores, boxes = im_detect(sess,net, im)
    timer.toc()
 
    print('Detection took {:.3f}s for '
          '{:d} object proposals'.format(timer.total_time, boxes.shape[0]))
    frameRate = 1.0 / timer.total_time
    print('fps:'+str(float(frameRate)))
 
    # Visualize detections for each class
    CONF_THRESH = 0.8
    NMS_THRESH = 0.3
    for cls_ind, cls in enumerate(CLASSES[1:]):
        # because we skipped background
        cls_ind += 1
        cls_boxes = boxes[:, 4 * cls_ind:4 * (cls_ind + 1)]
        cls_scores = scores[:, cls_ind]
        dets = np.hstack((cls_boxes,
                          cls_scores[:, np.newaxis])).astype(np.float32)
        keep = nms(dets, NMS_THRESH)
        dets = dets[keep, :]
        vis_detections_video(im, cls, dets, thresh=CONF_THRESH)
 
        text='{:s} {:.2f}'.format("FPS:", frameRate)
        position=(50, 50)
        cv2.putText(im, text, position, cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255))
        cv2.imshow(videoFilePath.split('/')[len(videoFilePath.split('/')) - 1], im)
        cv2.waitKey(50)
 
def parse_args():
    """Parse input arguments."""
    parser = argparse.ArgumentParser(description='Tensorflow Faster R-CNN demo')
    parser.add_argument('--net', dest='demo_net', help='Network to use [vgg16 res101]',
                        choices=NETS.keys(), default='vgg16')
    parser.add_argument('--dataset', dest='dataset', help='Trained dataset [pascal_voc pascal_voc_0712]',
                        choices=DATASETS.keys(), default='pascal_voc')
    args = parser.parse_args()
 
    return args
 
if __name__ == '__main__':
    args = parse_args()
 
    # model path
    demonet = args.demo_net
    dataset = args.dataset
    tfmodel = os.path.join('output', demonet, DATASETS[dataset][0], 'default', NETS[demonet][0])
 
    if not os.path.isfile(tfmodel + '.meta'):
        print(tfmodel)
        raise IOError(('{:s} not found.\nDid you download the proper networks from '
                       'our server and place them properly?').format(tfmodel + '.meta'))
 
    # set config
    tfconfig = tf.ConfigProto(allow_soft_placement=True)
    tfconfig.gpu_options.allow_growth = True
 
    # load network
    if demonet == 'vgg16':
        net = vgg16(batch_size=1)
 
    else:
        raise NotImplementedError
 
    # init session
    sess = tf.Session(config=tfconfig)
 
    net.create_architecture(sess, "TEST", 21,
                            tag='default', anchor_scales=[8, 16, 32])
    saver = tf.train.Saver()
    saver.restore(sess, tfmodel)
 
    print('\n\nLoaded network {:s}'.format(tfmodel))
 
    # Warmup on a dummy image
    im = 128 * np.ones((300, 500, 3), dtype=np.uint8)
    for i in range(2):
        _, _ = im_detect(sess,net, im)
 
    videoFilePath = 'Camera Road 01.avi'
    videoCapture = cv2.VideoCapture(videoFilePath)
 
    while True:
        success, im = videoCapture.read()
        demo(net, im)
 
        if cv2.waitKey(10) & 0xFF == ord('q'):
            break
    videoCapture.release()
    cv2.destroyAllWindows()
    sess.close()

码农公寓

一、环境配置

二、准备数据集

三、模型下载

四、 训练模型

五、测试模型

填坑

相关文章

四、训练模型