传统目标检测实战：HOG+SVM

文章目录

• 传统目标检测实战：HOG+SVM
• • 1. 前言
  • • 1.1 传统和深度
    • 1.2 何为传统目标检测
    • 1.3 传统目标检测方法不足
  • 2. 先验知识
  • 3. 项目框架
  • • 3.1 文件架构
    • 3.2 方法简要介绍
  • 4. 工具函数（utils.py）
  • 5. 特征提取（extract_feature.py）
  • 6. 训练分类器（train.py）
  • 7. 测试（test.py）
  • 8. 困难样本挖掘（neg_mining.py）
  • 9. 总结

1. 前言

1.1 传统和深度

在深度学习出现之前，传统的目标检测方法大概分为区域选择（滑窗）、特征提取（SIFT、HOG等）、**分类器（SVM、Adaboost等）**三个部分，其主要问题有两方面：一方面滑窗选择策略没有针对性、时间复杂度高，窗口冗余；另一方面手工设计的特征鲁棒性较差。自深度学习出现之后，目标检测取得了巨大的突破，最瞩目的两个方向有：

• 以RCNN为代表的基于Region Proposal的深度学习目标检测算法（RCNN，SPP-NET，Fast-RCNN，Faster-RCNN等），它们是two-stage的，需要先使用启发式方法（selective search）或者CNN网络（RPN）产生Region Proposal，然后再在Region Proposal上做分类与回归。
• 以YOLO为代表的基于回归方法的深度学习目标检测算法（YOLO，SSD等）,其仅仅使用一个CNN网络直接预测不同目标的类别与位置。

1.2 何为传统目标检测

首先我们先来了解一下什么是目标检测？简单来说就是把存在的目标从图片中找到并识别出来。我们发现这对于我们人来说十分简单，但对于计算机而言，它是怎么做到的呢？

• 传统目标检测方法分为三部分：区域选择 → 特征提取 → 分类器
  即首先在给定的图像上选择一些候选的区域，然后对这些区域提取特征，最后使用训练的分类器进行分类。下面我们对这三个阶段分别进行介绍。

1. 区域选择：这一步是为了对目标的位置进行定位。由于目标可能出现在图像的任何位置，而且目标的大小、长宽比例也不确定，所以最初采用滑动窗口的策略对整幅图像进行遍历，而且需要设置不同的尺度，不同的长宽比。这种穷举的策略虽然包含了目标所有可能出现的位置，但是缺点也是显而易见的：时间复杂度太高，产生冗余窗口太多，这也严重影响后续特征提取和分类的速度和性能。（实际上由于受到时间复杂度的问题，滑动窗口的长宽比一般都是固定的设置几个，所以对于长宽比浮动较大的多类别目标检测，即便是滑动窗口遍历也不能得到很好的区域）。

2. 特征提取：由于目标的形态多样性，光照变化多样性，背景多样性等因素使得设计一个鲁棒的特征并不是那么容易。然而提取特征的好坏直接影响到分类的准确性。（这个阶段常用的特征有SIFT、HOG等）

3. 分类器：主要有SVM，Adaboost等。

1.3 传统目标检测方法不足

总结一下，传统目标检测存在的两个主要问题：

• 基于滑动窗口的区域选择策略没有针对性，时间复杂度高，窗口冗余；
• 手工设计的特征对于多样性的变化并没有很好的鲁棒性。

2. 先验知识

1. 机器视觉特征简单介绍：HOG、SIFT、SURF、ORB、LBP、HAAR
2. skimage.feature–corner_harris、hog、local_binary_pattern说明
3. SVM很好理解，就是分类器

3. 项目框架

**本文主要是针对裂缝目标检测的项目。**也可扩展到其他的目标检测。！！！

3.1 文件架构

• data代表数据文件夹，下有img（原始图像有正样本（有裂缝）和负样本（无裂缝））、feature（对原始图像提取特征后，得到正样本特征和负样本特征）、model（保存训练好的分类器）、result（预测的结果）和test_img（预测时所用到的图像）。
• 其余的py文件，下面详细介绍。

3.2 方法简要介绍

4. 工具函数（utils.py）

from skimage.feature import hog, local_binary_pattern# settings for LBP
radius = 3
n_points = 8 * radius# settings for HOG
ppc = (32, 32)
cpb = (3, 3)def get_hog_feature(img):return hog(img, orientations=9, pixels_per_cell=ppc, cells_per_block=cpb, block_norm='L1', transform_sqrt=False, feature_vector=True)def get_lbp_feature(img):return local_binary_pattern(img, n_points, radius)def sliding_window(image, window_size, step_size):for row in range(0, image.shape[0], step_size[0]):for col in range(0, image.shape[1], step_size[1]):yield (row, col, image[row:row + window_size[0], col:col + window_size[1]])def overlapping_area(detection_1, detection_2, show = False):'''计算两个检测区域覆盖大小，detection：(x, y, pred_prob, width, height, area)'''# Calculate the x-y co-ordinates of the# rectangles# detection_1的 top left 和 bottom rightx1_tl = detection_1[0]y1_tl = detection_1[1]x1_br = detection_1[0] + detection_1[3]y1_br = detection_1[1] + detection_1[4]# detection_2的 top left 和 bottom rightx2_tl = detection_2[0]y2_tl = detection_2[1]x2_br = detection_2[0] + detection_2[3]y2_br = detection_2[1] + detection_2[4]# 计算重叠区域x_overlap = max(0, min(x1_br, x2_br) - max(x1_tl, x2_tl))y_overlap = max(0, min(y1_br, y2_br) - max(y1_tl, y2_tl))overlap_area = x_overlap * y_overlaparea_1 = detection_1[3] * detection_1[4]area_2 = detection_2[3] * detection_2[4]# 1. 重叠比例计算1# total_area = area_1 + area_2 - overlap_area# return overlap_area / float(total_area)# 2.重叠比例计算2area = area_1if area_1 < area_2: area = area_2return float(overlap_area / area)def nms(detections, threshold=0.5):'''抑制策略：1. 最大的置信值先行2. 最大的面积先行非极大值抑制减少重叠区域, detection:(x, y, pred_prob, width, height)'''if len(detections) == 0:return []# Sort the detections based on confidence score# 根据预测值大小排序预测结果detections = sorted(detections, key=lambda detections: detections[2], reverse=True)# print((detections[0][5], detections[-1][5]))# Unique detections will be appended to this list# 非极大值抑制后的检测区域new_detections=[]# Append the first detection# 默认第一个区域置信度最高是正确检测区域new_detections.append(detections[0])# Remove the detection from the original list# 去除以检测为正确的区域del detections[0]# For each detection, calculate the overlapping area# and if area of overlap is less than the threshold set# for the detections in `new_detections`, append the# detection to `new_detections`.# In either case, remove the detection from `detections` list.print(len(detections))for index, detection in enumerate(detections):if len(new_detections) >= 20:breakoverlapping_small = True# 重叠区域过大，则删除该区域，同时结束检测，过小则继续检测for new_detection in new_detections:if overlapping_area(detection, new_detection) > threshold:overlapping_small = Falsebreak# 整个循环中重叠区域都小那么增加if overlapping_small:new_detections.append(detection)return new_detections

5. 特征提取（extract_feature.py）

import numpy as np
import joblib
import os
import glob
from utils import *
import cv2# setting for img_resize
window_size = (256, 256)train_dataset_path = os.path.expanduser('./data/img')
feat_path = './data/feature'
feat_pos_path = os.path.join(feat_path, 'pos')
feat_neg_path = os.path.join(feat_path, 'neg')train_dataset_pos_lists = glob.glob(os.path.join(train_dataset_path, 'pos/*.jpg'))
train_dataset_neg_lists = glob.glob(os.path.join(train_dataset_path, 'neg/*.jpg'))# 正样本特征存储
for pos_path in train_dataset_pos_lists:pos_im = cv2.imread(pos_path, cv2.IMREAD_GRAYSCALE)pos_im = cv2.resize(pos_im, window_size)pos_lbp = get_hog_feature(pos_im)pos_lbp = pos_lbp.reshape(-1)feat_pos_name = os.path.splitext(os.path.basename(pos_path))[0] + '.feat'joblib.dump(pos_lbp, os.path.join(feat_pos_path, feat_pos_name))# 负样本特征存储
for neg_path in train_dataset_neg_lists:neg_im = cv2.imread(neg_path, cv2.IMREAD_GRAYSCALE)neg_im = cv2.resize(neg_im, window_size)neg_lbp = get_hog_feature(neg_im)neg_lbp = neg_lbp.reshape(-1)feat_neg_name = os.path.splitext(os.path.basename(neg_path))[0] + '.feat'joblib.dump(neg_lbp, os.path.join(feat_neg_path, feat_neg_name))

6. 训练分类器（train.py）

from sklearn.svm import LinearSVC, SVC
import numpy as np
import joblib
import os
import globfeat_path = './data/feature'
feat_pos_path = os.path.join(feat_path, 'pos')
feat_neg_path = os.path.join(feat_path, 'neg')
train_feat_pos_lists = glob.glob(os.path.join(feat_pos_path, '*.feat'))
train_feat_neg_lists = glob.glob(os.path.join(feat_neg_path, '*.feat'))
X = []
y = []# 加载正例样本
for feat_pos in train_feat_pos_lists:feat_pos_data = joblib.load(feat_pos)X.append(feat_pos_data)y.append(1)
#     print('feat_pos_data.shape:', feat_pos_data.shape)# 加载负例样本
for feat_neg in train_feat_neg_lists:feat_neg_data = joblib.load(feat_neg)X.append(feat_neg_data)y.append(0)
#     print('feat_neg_data.shape:', feat_neg_data.shape)clf = LinearSVC(dual = False)
# clf = SVC()
clf.fit(X, y)
clf.score(X, y)model_path = './data/model'
joblib.dump(clf, os.path.join(model_path, 'svm.model'))
print(len(X),X[0].shape)

7. 测试（test.py）

import numpy as np
import joblib
import os
import glob
from skimage.transform import pyramid_gaussian
import cv2
from utils import *window_size = (256, 256)step_size = (128, 128)img_name = '2.jpg'
test_image = cv2.imread("./data/test_img/" + img_name, cv2.IMREAD_GRAYSCALE)model_path = './data/model'
clf = joblib.load(os.path.join(model_path, 'svm.model'))scale = 0
detections = []
downscale = 1.25for test_image_pyramid in pyramid_gaussian(test_image, downscale=downscale):if test_image_pyramid.shape[0] < window_size[0] or test_image_pyramid.shape[1] < window_size[1]:breakfor (row, col, sliding_image) in sliding_window(test_image_pyramid, window_size, step_size):if sliding_image.shape != window_size:continuesliding_image_lbp = get_hog_feature(sliding_image)sliding_image_lbp = sliding_image_lbp.reshape(1, -1)pred = clf.predict(sliding_image_lbp)if pred==1:pred_prob = clf.decision_function(sliding_image_lbp)(window_height, window_width) = window_sizereal_height = int(window_height*downscale**scale)real_width = int(window_width*downscale**scale)detections.append((int(col*downscale**scale), int(row*downscale**scale), pred_prob, real_width, real_height, real_height*real_width))scale+=1test_image1 = cv2.imread("./data/test/" + img_name, 1)
test_image_detect = test_image1.copy()
for detection in detections:col = detection[0]row = detection[1]width = detection[3]height = detection[4]cv2.rectangle(test_image_detect, pt1=(col, row), pt2=(col+width, row+height), color=(255, 0, 0), thickness=4)print('before NMS')
cv2.imwrite("./data/result/_"+ img_name, test_image_detect)threshold = 0.2
detections_nms = nms(detections, threshold)test_image_detect = test_image1.copy()
for detection in detections_nms:col = detection[0]row = detection[1]width = detection[3]height = detection[4]cv2.rectangle(test_image_detect, pt1=(col, row), pt2=(col+width, row+height), color=(0, 255, 0), thickness=4)print('after NMS')
cv2.imwrite("./data/result/"+ img_name, test_image_detect)

8. 困难样本挖掘（neg_mining.py）

由于预测的结果存在很大的偏差，在于训练不到位，这时候需要将那些预测错误的样本再次训练。

from sklearn.svm import LinearSVC, SVC
import matplotlib.pyplot as plt
import numpy as np
from scipy import misc
import joblib
import os
import glob
from skimage.feature import hog
feat_path = './data/feature'
feat_pos_path = os.path.join(feat_path, 'pos')
feat_neg_path = os.path.join(feat_path, 'neg')
train_feat_pos_lists = glob.glob(os.path.join(feat_pos_path, '*.feat'))
train_feat_neg_lists = glob.glob(os.path.join(feat_neg_path, '*.feat'))model_path = './data/model'
clf = joblib.load(os.path.join(model_path, 'svm.model'))# 128
for i in range(0, 5):x = []y = []cnt = 0for feat_neg in train_feat_neg_lists:feat_neg_data = joblib.load(feat_neg)feat_neg_data1 = feat_neg_data.reshape(1, -1)pred = clf.predict(feat_neg_data1)if pred == 1:cnt += 1x.append(feat_neg_data)y.append(0)cnt1 = 0for feat_pos in train_feat_pos_lists:feat_pos_data = joblib.load(feat_pos)feat_pos_data1 = feat_pos_data.reshape(1, -1)pred = clf.predict(feat_pos_data1)if pred == 0:cnt1 += 1x.append(feat_pos_data)y.append(1)clf.fit(x, y)print(cnt, cnt1)# model_path = './data/model'
# joblib.dump(clf, os.path.join(model_path, 'svm.model'))

9. 总结

通过这样的目标检测代码框架，可以得出还不错的结果。有错误欢迎指出！