YOLOv2是Joseph Redmon提出的針對(duì)YOLO算法不足的改進(jìn)版本，作者使用了一系列的方法對(duì)原來的YOLO多目標(biāo)檢測(cè)框架進(jìn)行了改進(jìn)，在保持原有速度的優(yōu)勢(shì)之下，精度上得以提升，此外作者提出了一種目標(biāo)分類與檢測(cè)的聯(lián)合訓(xùn)練方法，通過這種方法YOLO9000可以同時(shí)在COCO和ImageNet數(shù)據(jù)集中進(jìn)行訓(xùn)練，訓(xùn)練后的模型可以實(shí)現(xiàn)多達(dá)9000種物體的實(shí)時(shí)檢測(cè)。

Paper：https://arxiv.org/abs/1612.08242
Github：https://github.com/pjreddie/darknet
Website：https://pjreddie.com/darknet/yolo

作者為YOLO算法設(shè)計(jì)了獨(dú)有的深度學(xué)習(xí)框架darknet，因此沒有提供Python的接口。在實(shí)驗(yàn)中，我找到了兩種在Python 3中使用YOLOv2網(wǎng)絡(luò)的方法。

第一種：為darknet添加Python接口

Github：https://github.com/SidHard/py-yolo2

該項(xiàng)目使用了原始的darknet網(wǎng)絡(luò)，需要使用cmake重新編譯源碼，因此在Linux上使用更為方便一些。

首先從git上下載該項(xiàng)目
git clone https://github.com/SidHard/py-yolo2.git
執(zhí)行cmake生成項(xiàng)目
cmake .. && make
最后執(zhí)行yolo.py測(cè)試項(xiàng)目，相應(yīng)的網(wǎng)絡(luò)結(jié)構(gòu).cfg文件保存在cfg文件夾中，權(quán)值.weight文件放在根目錄下，這些可以從darknet的官方網(wǎng)站上下載使用。

第二種：使用keras

Github：https://github.com/allanzelener/YAD2K

該項(xiàng)目使用了keras與tensorflow-gpu，因此可以在任何使用該框架的環(huán)境下運(yùn)行，我在自己的程序中使用的該種方法。

首先下載源文件并且配置環(huán)境，可以使用anaconda環(huán)境或者在全局安裝。

git clone https://github.com/allanzelener/yad2k.git
cd yad2k

# [Option 1] To replicate the conda environment:
conda env create -f environment.yml
source activate yad2k
# [Option 2] Install everything globaly.
pip install numpy

pip install tensorflow-gpu  # CPU-only: conda install -c conda-forge tensorflow
pip install keras # Possibly older release: conda install keras

快速開始

• 從Darknet官方下載model：official YOLO website.
  wget http://pjreddie.com/media/files/yolo.weights
• 將 Darknet YOLO_v2 model轉(zhuǎn)換為Keras model.
  ./yad2k.py cfg/yolo.cfg yolo.weights model_data/yolo.h5
• 測(cè)試圖片位于 images/文件夾.
  ./test_yolo.py model_data/yolo.h5

最后執(zhí)行test_yolo就可以執(zhí)行網(wǎng)絡(luò)，在images/out/文件夾里可以看到執(zhí)行效果。

dog.jpg

eagle.jpg

giraffe.jpg

horses.jpg

為了方便模型用于測(cè)試視頻與圖片，我對(duì)demo做了修改，相比原來的測(cè)試代碼，能夠直接移植到項(xiàng)目中去，對(duì)象化的程序也更易于修改，代碼如下

#! /usr/bin/env python
"""Run a YOLO_v2 style detection model on test images."""
import cv2
import os
import time
import numpy as np
from keras import backend as K
from keras.models import load_model

from yad2k.models.keras_yolo import yolo_eval, yolo_head


class YOLO(object):
    def __init__(self):
        self.model_path = 'model_data/yolo.h5'
        self.anchors_path = 'model_data/yolo_anchors.txt'
        self.classes_path = 'model_data/coco_classes.txt'
        self.score = 0.3
        self.iou = 0.5

        self.class_names = self._get_class()
        self.anchors = self._get_anchors()
        self.sess = K.get_session()
        self.boxes, self.scores, self.classes = self.generate()

    def _get_class(self):
        classes_path = os.path.expanduser(self.classes_path)
        with open(classes_path) as f:
            class_names = f.readlines()
        class_names = [c.strip() for c in class_names]
        return class_names

    def _get_anchors(self):
        anchors_path = os.path.expanduser(self.anchors_path)
        with open(anchors_path) as f:
            anchors = f.readline()
            anchors = [float(x) for x in anchors.split(',')]
            anchors = np.array(anchors).reshape(-1, 2)
        return anchors

    def generate(self):
        model_path = os.path.expanduser(self.model_path)
        assert model_path.endswith('.h5'), 'Keras model must be a .h5 file.'

        self.yolo_model = load_model(model_path)

        # Verify model, anchors, and classes are compatible
        num_classes = len(self.class_names)
        num_anchors = len(self.anchors)
        # TODO: Assumes dim ordering is channel last
        model_output_channels = self.yolo_model.layers[-1].output_shape[-1]
        assert model_output_channels == num_anchors * (num_classes + 5), \
            'Mismatch between model and given anchor and class sizes'
        print('{} model, anchors, and classes loaded.'.format(model_path))

        # Check if model is fully convolutional, assuming channel last order.
        self.model_image_size = self.yolo_model.layers[0].input_shape[1:3]
        self.is_fixed_size = self.model_image_size != (None, None)

        # Generate output tensor targets for filtered bounding boxes.
        # TODO: Wrap these backend operations with Keras layers.
        yolo_outputs = yolo_head(self.yolo_model.output, self.anchors, len(self.class_names))
        self.input_image_shape = K.placeholder(shape=(2, ))
        boxes, scores, classes = yolo_eval(yolo_outputs, self.input_image_shape, score_threshold=self.score, iou_threshold=self.iou)
        return boxes, scores, classes

    def detect_image(self, image):
        start = time.time()
        y, x, _ = image.shape

        if self.is_fixed_size:  # TODO: When resizing we can use minibatch input.
            resized_image = cv2.resize(image, tuple(reversed(self.model_image_size)), interpolation=cv2.INTER_CUBIC)
            image_data = np.array(resized_image, dtype='float32')
        else:
            image_data = np.array(image, dtype='float32')

        image_data /= 255.
        image_data = np.expand_dims(image_data, 0)  # Add batch dimension.

        out_boxes, out_scores, out_classes = self.sess.run(
            [self.boxes, self.scores, self.classes],
            feed_dict={
                self.yolo_model.input: image_data,
                self.input_image_shape: [image.shape[0], image.shape[1]],
                K.learning_phase(): 0
            })
        print('Found {} boxes for {}'.format(len(out_boxes), 'img'))

        for i, c in reversed(list(enumerate(out_classes))):
            predicted_class = self.class_names[c]
            box = out_boxes[i]
            score = out_scores[i]

            label = '{} {:.2f}'.format(predicted_class, score)
            top, left, bottom, right = box
            top = max(0, np.floor(top + 0.5).astype('int32'))
            left = max(0, np.floor(left + 0.5).astype('int32'))
            bottom = min(y, np.floor(bottom + 0.5).astype('int32'))
            right = min(x, np.floor(right + 0.5).astype('int32'))
            print(label, (left, top), (right, bottom))

            cv2.rectangle(image, (left, top), (right, bottom), (255, 0, 0), 2)
            cv2.putText(image, label, (left, int(top - 4)), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA)
        end = time.time()
        print(end - start)
        return image

    def close_session(self):
        self.sess.close()


def detect_vedio(video, yolo):
    camera = cv2.VideoCapture(video)
    cv2.namedWindow("detection", cv2.WINDOW_NORMAL)
    while True:
        res, frame = camera.read()

        if not res:
            break

        image = yolo.detect_image(frame)
        cv2.imshow("detection", image)
        if cv2.waitKey(110) & 0xff == 27:
                break
    yolo.close_session()


def detect_img(img, yolo):
    image = cv2.imread(img)
    r_image = yolo.detect_image(image)
    cv2.namedWindow("detection")
    while True:
        cv2.imshow("detection", r_image)
        if cv2.waitKey(110) & 0xff == 27:
                break
    yolo.close_session()


if __name__ == '__main__':
    yolo = YOLO()
    img = 'E:\Documents\Downloads\YAD2K-master\YAD2K-master\images\horses.jpg'
    video = 'E:\Documents\Documents\python\Traffic\data\person.avi'  
    detect_img(img, yolo)
    detect_vedio(video, yolo)