OpenCV中训练CascadeClassifier（级联分类器）

参考:

---

Cascade Classifier Training
[OpenCV3]级联分类器训练——traincascade快速使用详解

级联分类器介绍

---

• 文章总概：Cascade分类器全程级联增强弱分类器(boosted cascade of weak classifiers)，而使用此分类器主要包括两方面------训练&定位。而麻烦的在于训练，如何训练，正是这篇文章将要介绍的：1.收集数据，2.准备训练数据以及训练模型。
• 个人理解：人脸检测中一个比较好的算法是mtcnn网络，也是一个级联分类器，与Cascade的区别在于一个用的卷积神经网络，一个用的是Haar或者LBP提取特征。而弱分类器的原理可以总结为：一个分类器通过学习而能够分辨的特征的能力有限，但如果让多个分类器去学习识别不同的特征，这样通过多个分类器进行分类后，所产生的结果错误率将极大降低。

数据准备

---

不更新了，只能用CPU训练的速度太慢了，而且网上都说这种方法的正确率不太好，所以不更新了

• 可以通过官方给的样本做正样本标签，也可以通过此工具labelImg，建议使用后者，后者在大数据集方面使用较多(列如微软VOC数据集)。

正样本以及负样本生成

---

正负样本生成脚本

import sys
import numpy as np
import xml.etree.ElementTree as ET
import cv2
import os
import numpy.random as npr
from utils import IoU
from utils import ensure_directory_exists

save_dir = "/home/rui"
anno_path = "./firepos/annotation"
im_dir = "./firepos/images"
pos_save_dir = os.path.join(save_dir, "./res/positive")
neg_save_dir = os.path.join(save_dir, './res/negative')

ensure_directory_exists(pos_save_dir)
ensure_directory_exists(neg_save_dir)


names_xml = os.listdir(anno_path)

img_rule_h = 45
img_rule_w = 45

size = img_rule_h

num = len(names_xml)
print "%d pics in total" % num
p_idx = 0 # positive
n_idx = 0 # negative
d_idx = 0 # dont care
idx = 0
box_idx = 0
for ne_xml in names_xml:
    tree = ET.parse(os.path.join(anno_path, ne_xml))
    root = tree.getroot()
    loc_bbox = []
    width_xml = root.find("size").find("width").text
    height_xml = root.find("size").find("height").text
    for node in root.findall('object'):
        label_ = node.find('name').text
        if label_ == "fire":
            xmin_ = node.find('bndbox').find('xmin').text
            ymin_ = node.find('bndbox').find('ymin').text
            xmax_ = node.find('bndbox').find('xmax').text
            ymax_ = node.find('bndbox').find('ymax').text
            loc_bbox.append(xmin_)
            loc_bbox.append(ymin_)
            loc_bbox.append(xmax_)
            loc_bbox.append(ymax_)
    im_path = "{}/{}".format(im_dir, ne_xml.split(".")[0])
    if os.path.exists(im_path + ".jpg"):
        im_path = "{}.jpg".format(im_path)
    else:
        im_path = "{}.JPG".format(im_path)
    boxes = np.array(loc_bbox, dtype=np.float32).reshape(-1, 4)
    img = cv2.imread(im_path)
    h, w, c =img.shape
    if h != int(height_xml) or w != int(width_xml):
        print h, height_xml,w,width_xml
        continue
    idx += 1
    if idx % 100 == 0:
        print idx, "images done"

    height, width, channel = img.shape

    neg_num = 0
    while neg_num < 700:
        size_new = 0.0
        if width > height:
            size_new = npr.randint(img_rule_h + 1, max(img_rule_h, height / 2 - 1))
        else:
            size_new = npr.randint(img_rule_w + 1, max(img_rule_w, width / 2 - 1))

        size_new = int(size_new)
        nx = npr.randint(0, width - size_new)
        ny = npr.randint(0, height - size_new)
        crop_box = np.array([nx, ny, nx + size_new, ny + size_new])
        Iou = IoU(crop_box, boxes)
        cropped_im = img[ny : ny + size_new, nx : nx + size_new, :]
        resized_im = cv2.resize(cropped_im, (img_rule_w, img_rule_h), interpolation=cv2.INTER_LINEAR)

        if len(Iou) != 0:
            if np.max(Iou) < 0.1:
            # Iou with all gts must below 0.3
                save_file = os.path.join(neg_save_dir, "%s.jpg"%n_idx)
                cv2.imwrite(save_file, resized_im)
                n_idx += 1
                neg_num += 1
        else:
            # Iou with all gts must below 0.3

            save_file = os.path.join(neg_save_dir, "%s.jpg"%n_idx)
            cv2.imwrite(save_file, resized_im)
            n_idx += 1
            neg_num += 1

    for box in boxes:
        # box (x_left, y_top, x_right, y_bottom)
        x1, y1, x2, y2 = box
        w = x2 - x1 + 1
        h = y2 - y1 + 1

#        if float(w) / h < 2:
#            continue

        # ignore small faces
        # in case the ground truth boxes of small faces are not accurate
        if w < img_rule_w or h < img_rule_h or x1 < 0 or y1 < 0:
            continue

        # generate positive examples and part faces
        pos_nums = 300
        while pos_nums > 0:
            size_new = npr.randint(int(pow(w * h, 0.5) - 1), int(max(w, h)))
            # delta here is the offset of box center

            delta_x = npr.randint(int(-size_new * 0.1), int(size_new * 0.1))
            delta_y = npr.randint(int(-size_new * 0.1), int(size_new * 0.1))

            nx1 = max(x1 + w / 2 + delta_x - size_new / 2, 0)
            ny1 = max(y1 + h / 2 + delta_y - size_new / 2, 0)
            nx2 = min(width, nx1 + size_new)
            ny2 = min(height, ny1 + size_new)
            if nx2 > width or ny2 > height:
                continue
            crop_box = np.array([nx1, ny1, nx2, ny2])

            cropped_im = img[int(ny1) : int(ny2), int(nx1) : int(nx2), :]
            resized_im = cv2.resize(cropped_im, (img_rule_w, img_rule_h))

            box_ = box.reshape(1, -1)

            pos_nums -= 1
            save_file = os.path.join(pos_save_dir, "%s.jpg"%p_idx)
            cv2.imwrite(save_file, resized_im)
            p_idx += 1
        box_idx += 1
        print "%s images done, pos: %s, neg: %s"%(idx, p_idx, n_idx)

将正样本写入

import os

pos_dir = "/home/rui/res/positive"

pos_list = os.listdir(pos_dir)

f = open("/home/rui/temp.txt", "w")

for im in pos_list:
    name = "positive/{} 1 0 0 45 45\n".format(im)
    print name
    f.writelines(name)
f.close()

find -name *.jpg >> neg.txt

待更新