tensorflow使用object detection实现目标检测超详细全流程(视频+图像集检测)

谁践踏了优雅 2022-04-05 18:11 336阅读 0赞

使用tensorflow object detection进行训练检测。参考原始代码：[https://github.com/tensorflow/models/tree/master/research][https_github.com_tensorflow_models_tree_master_research]  
本博客以mobilenet-ssd-v2为例进行处理，通过换模型即可实现faster RCNN等的训练检测。

## 1、数据整理 ##

对生成的数据集（整理成VOC格式），通过Annotations的数据数进行train、test、val、trainval.txt的生成  
进入目录

cd VOCdevkit/VOC2012/
    python data_segment.py

""" data_segment.py 可自主设计数据集的比例，即trainval_percent,train_percent """
    import os  
    import random  
    
    trainval_percent = 0.9  
    train_percent = 0.95  
    xmlfilepath = 'Annotations'  
    txtsavepath = 'ImageSets\Main'  
    total_xml = os.listdir(xmlfilepath)  
    
    num=len(total_xml)  
    list1=range(num)  
    tv=int(num*trainval_percent)  
    tr=int(tv*train_percent)  
    trainval= random.sample(list1,tv)  
    train=random.sample(trainval,tr) 
    ftrainval = open('ImageSets/Main/trainval.txt', 'w')  
    ftest = open('ImageSets/Main/test.txt', 'w')  
    ftrain = open('ImageSets/Main/train.txt', 'w')  
    fval = open('ImageSets/Main/val.txt', 'w')  
    for i in list1:  
        name=total_xml[i][:-4]+'\n'  
        if i in trainval:  
            ftrainval.write(name)  
            if i in train:  
                ftrain.write(name)  
            else:  
                fval.write(name)  
        else:  
            ftest.write(name)  
    
    ftrainval.close()  
    ftrain.close()  
    fval.close()  
    ftest .close()

#### note： ####

在使用数据集格式转化工具，生成的voc文件中，ImageSets/mains中，只含有trainval.txt。（可通过上述方式进行重新生成.txt文件，亦可暂时忽略，在后续生成record文件时，直接应用）  
数据集的下载可参考[自动驾驶数据集][Link 1]，同时能获得自动驾驶数据集与voc格式之间的转换。

## 2、安装依赖项 ##

pip install -r requirements.txt

## 3、载入环境变量 ##

protoc object_detection/protos/*.proto --python_out=.
    export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/object_detection
    export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

## 4、生成record文件 ##

# 生成相应的train.record、val.record、test.record
    python object_detection/dataset_tools/create_pascal_tf_record.py --data_dir=VOCdevkit/ --year=VOC2012 --set=val --label_map_path=object_detection/data/pascal_label_map.pbtxt --output_path=dataset/val.record

也可根据修改过的record生成文件，直接对数据集转化结果trainval.txt进行处理。

python object_detection/create_crj.py --data_dir=VOCdevkit/ --set=trainval --year=VOC2012 --label_map_path=object_detection/data/pascal_label_map.pbtxt

""" 前面部分，篇幅问题，省略，参考create_pascal_tf_record.py中函数 """
    def main(_):
        if FLAGS.set not in SETS:
            raise ValueError('set must be in : {}'.format(SETS))
        if FLAGS.year not in YEARS:
            raise ValueError('year must be in : {}'.format(YEARS))
    
        data_dir = FLAGS.data_dir
        years = ['VOC2007', 'VOC2012']
        if FLAGS.year != 'merged':
            years = [FLAGS.year]
        # writer = tf.python_io.TFRecordWriter(FLAGS.output_path)
        label_map_dict = label_map_util.get_label_map_dict(FLAGS.label_map_path)
        for year in years:
            logging.info('Reading from PASCAL %s dataset.', year)
            # examples_path = os.path.join(data_dir, year, 'ImageSets', 'Main',
            # 'aeroplane_' + FLAGS.set + '.txt')
            examples_path = os.path.join(data_dir, year, 'ImageSets', 'Main', FLAGS.set + '.txt')
            image_dir = os.path.join(data_dir, year, 'JPEGImages')
            annotations_dir = os.path.join(data_dir, year, FLAGS.annotations_dir)
            examples_list = dataset_util.read_examples_list(examples_path)
            trainval_percent = 0.8
            train_percent = 0.7
            
            random.seed(42)
            random.shuffle(examples_list)
            num_examples = len(examples_list)
            num_trainval = int(trainval_percent * num_examples)
            num_train = int(train_percent * num_trainval)
            trainval_examples = examples_list[:num_trainval]
            train_examples = trainval_examples[:num_train]
            val_examples = trainval_examples[num_train:]
            test_examples = examples_list[num_trainval:]
            logging.info('%d training and %d validation examples.', len(train_examples), len(val_examples))
            train_output_path = 'record_path/train.record'
            val_output_path = 'record_path/val.record'
            test_output_path = 'record_path/test.record'
            create_pascal_record(train_output_path, label_map_dict,annotations_dir, image_dir, train_examples)
            create_pascal_record(val_output_path, label_map_dict,annotations_dir, image_dir, val_examples)
            create_pascal_record(test_output_path, label_map_dict,annotations_dir, image_dir, test_examples)
    
    if __name__ == '__main__':
        tf.app.run()

## 5、训练 ##

#### 修改ssd\_mobilenet\_v2\_coco.config中的六处 ####

（1）num\_classes: 3（**设为自己的训练集的类别**）  
（2）fine\_tune\_checkpoint: “/home/crj/tensorflow/ssd\_mobilenet\_v2\_coco\_2018\_03\_29/model.ckpt”（**fine\_tune\_checkpoint的地址**）  
（3）train\_input\_reader：\{\}中input\_path和label\_map\_path的**路径**  
（4）eval\_input\_reader：\{\}中input\_path和label\_map\_path的**路径**  
（5）batch\_size：48 **通常不能太小，否则会报错**  
（6）initial\_learning\_rate: 0.005 **不能太小，否则容易陷入局部过拟合**

# 训练模型在dataset中保存
    python object_detection/legacy/train.py --logtostderr --train_dir=dataset --pipeline_config_path=ssd_mobilenet_v2_coco.config

## 6、导出训练数据 ##

将dataset中的生成模型导出到output\_inference\_graph输出文件夹中

python object_detection/export_inference_graph.py --input_type image_tensor --pipeline_config_path ssd_mobilenet_v2_coco.config --trained_checkpoint_prefix /home/crj/tensorflow/dataset/model.ckpt-20649 --output_directory output_inference_graph

![导出模型][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTA4MDE5OTQ_size_16_color_FFFFFF_t_70]

## 7、检测结果输出 ##

python object_detection/Object_Detection_Tensorflow_API_crj.py

from moviepy.editor import VideoFileClip
    from IPython.display import HTML
    # import detection_demo.object_detection_model as model
    import matplotlib
    matplotlib.use('Agg')
    import os
    import numpy as np
    import imageio
    imageio.plugins.ffmpeg.download()
    from object_detection.utils import label_map_util
    from object_detection.utils import visualization_utils as vis_util
    from moviepy.editor import VideoFileClip
    from IPython.display import HTML
    import time
    import tensorflow as tf
    import imageio
    imageio.plugins.ffmpeg.download()
    
    CWD_PATH = os.getcwd()
    # Path to frozen detection graph. This is the actual model that is used for the object detection.
    MODEL_NAME = 'output_inference_graph'
    # PATH_TO_CKPT = os.path.join(CWD_PATH, 'object_detection', MODEL_NAME, 'frozen_inference_graph.pb')
    PATH_TO_CKPT = os.path.join(CWD_PATH, '/home/crj/tensorflow', 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, '/home/crj/tensorflow/object_detection/data/pascal_label_map.pbtxt')
    
    NUM_CLASSES = 3
    
    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 process_image(image):
        # NOTE: The output you return should be a color image (3 channel) for processing video below
        # you should return the final output (image with lines are drawn on lanes)
        with detection_graph.as_default():
            with tf.Session(graph=detection_graph) as sess:
                image_process = detect_objects(image, sess, detection_graph)
                return image_process
    
    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
    
    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='')
            
    def process_image(image):
        # NOTE: The output you return should be a color image (3 channel) for processing video below
        # you should return the final output (image with lines are drawn on lanes)
        with detection_graph.as_default():
            with tf.Session(graph=detection_graph) as sess:
                image_process = detect_objects(image, sess, detection_graph)
                return image_process
    """ ================================视频检测=================================== """
    white_output = 'video1_out.mp4'
    clip1 = VideoFileClip("/home/crj/tensorflow/test_images1/1.mp4").subclip(0,2)
    white_clip = clip1.fl_image(process_image) #NOTE: this function expects color images!!s
    #get_ipython().magic('time white_clip.write_videofile(white_output, audio=False)')
    white_clip.write_videofile(white_output, audio=False)
    
    HTML(""" <video width="960" height="540" controls> <source src="{0}"> </video> """.format(white_output))
    """ ==============================图像检测===================================== """
    # First test on images
    PATH_TO_TEST_IMAGES_DIR = '/home/crj/tensorflow/test_images/'
    TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image-{}.jpg'.format(i)) for i in range(1, 5) ]
    # Size, in inches, of the output images.
    IMAGE_SIZE = (12, 8)
    
    def load_image_into_numpy_array(image):
      (im_width, im_height) = image.size
      return np.array(image.getdata()).reshape(
          (im_height, im_width, 3)).astype(np.uint8)
    
    from PIL import Image
    for image_path in TEST_IMAGE_PATHS:
        image = Image.open(image_path)
        image_np = load_image_into_numpy_array(image)
        plt.imshow(image_np)
        plt.show()
        print(image.size, image_np.shape)
        
    #Load a frozen TF model 
    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='')
    
    with detection_graph.as_default():
        with tf.Session(graph=detection_graph) as sess:
            for image_path in TEST_IMAGE_PATHS:
                image = Image.open(image_path)
                image_np = load_image_into_numpy_array(image)
                image_process = detect_objects(image_np, sess, detection_graph)
                print(image_process.shape)
                plt.figure(figsize=IMAGE_SIZE)
                plt.imshow(image_process)
                plt.show()

生成video1\_out.mp4文件  
![在这里插入图片描述][20181213163956460.png]  
实验结果如下  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTA4MDE5OTQ_size_16_color_FFFFFF_t_70 1]

**学习小记录**：  
（1）batchsize：批大小。在深度学习中，一般采用SGD训练，即每次训练在训练集中取batchsize个样本训练；  
（2）iteration：1个iteration等于使用batchsize个样本训练一次；  
（3）epoch：1个epoch等于使用训练集中的全部样本训练一次；  
举个例子，训练集有1000个样本，batchsize=10，那么：  
训练完整个样本集需要：100次iteration，1次epoch。

[https_github.com_tensorflow_models_tree_master_research]: https://github.com/tensorflow/models/tree/master/research
[Link 1]: https://blog.csdn.net/u010801994/article/details/84636952
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTA4MDE5OTQ_size_16_color_FFFFFF_t_70]: /images/20220405/d37e42ed1e4047f9bf62dba50c0bcf80.png
[20181213163956460.png]: /images/20220405/3d419f1fd9e548ac85e8a506768caa29.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3UwMTA4MDE5OTQ_size_16_color_FFFFFF_t_70 1]: /images/20220405/3e6451845e4c4faa99daf9067ad47fdc.png