【opencv】传统图像识别：hog+svm行人识别实战

# 加载本地图像数据集，并将数据集全部添加到列表中，然后打乱数据顺序
def load_image(filename):
paths = glob.glob(filename)
persons, labels = [], []
for i in paths:
persons.append(cv2.imread(i))
labels.append(1)
random.seed(1)
random.shuffle(persons)
persons = np.array(persons)
return persons, labels

# 图像预处理，将输入图像灰度化、高斯模糊
def image_preprocessing(image):
cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
# image = cv2.resize(image, dsize=(32, 64))
image_preprocess = cv2.GaussianBlur(image, (3, 3), sigmaX=1, sigmaY=1)
return image_preprocess

# 构建HOG检测器
def get_hog():
winSize = (64, 128)
cellSize = (8, 8)
blockSize = (16, 16)
blockStride = (16, 16)
nbins = 9
signedGradient = True
derivAperture = 1 # 默认参数
winSigma = -1. # 默认参数
histogramNormType = 0 # 默认参数
L2HysThreshold = 0.2 # 默认参数
gammaCorrection = 1 # 默认参数
nlevels = 64 # 默认参数
hog = cv2.HOGDescriptor(winSize, blockSize, blockStride, cellSize, nbins, derivAperture, winSigma,
histogramNormType, L2HysThreshold, gammaCorrection, nlevels, signedGradient)
return hog

# 创建SVM模型并配置参数
def SVM_model():
model = cv2.ml.SVM_create()
model.setType(cv2.ml.SVM_ONE_CLASS)
model.setKernel(cv2.ml.SVM_POLY)
model.setC(1)
model.setNu(0.01)
model.setDegree(0.1)
model.setCoef0(0.5)
model.setGamma(0.6)
model.setTermCriteria((cv2.TERM_CRITERIA_MAX_ITER, int(1e2), 1e-5))
return model

# 训练模型
def SVM_train(model, samples, responses):
model.train(samples, cv2.ml.ROW_SAMPLE, responses)
return model

# 计算分类器准确率
def accuracy(model, data_train, labels_train):
retval, result = model.predict(data_train)
temp = (np.array(labels_train) == result).mean()
print(f'该模型的准确率是：{temp * 100}')

# 测试分类器
def image_predict(model, data_test, samples, labels_test):
retval, result = model.predict(samples)
counter = 0
for i in (labels_test == result.ravel()):
# 测试结果与实际结果不符合仅呈现红色通道
if not i:
data_test[counter][..., :2] = 0
counter += 1
h1 = data_test[0]
for i in data_test[1:12, ...]:
h1 = np.hstack((h1, i))
h2 = data_test[12]
for i in data_test[13:, ...]:
h2 = np.hstack((h2, i))
return np.vstack((h1, h2))

if __name__ == "__main__":
print('加载图片...')
datas, labels = load_image('image\\Pedestrian detection\\per*.ppm')
temp, data_test = np.split(datas, [900])

print('数据预处理...')
datas = list(map(image_preprocessing, datas))

print('提取训练数据的HOG特征向量...')
hog = get_hog()
hog_vector = list(map(hog.compute, datas))

print('将数据集分为两部分，900张用于训练，24张用于测试...')
data_train, temp = np.split(datas, [900])
labels_train, labels_test = np.split(np.array(labels), [900])
hog_vector_train, hog_vector_test = np.split(hog_vector, [900])

print('训练SVM模型...')
model = SVM_model()
model_svm = SVM_train(model, hog_vector_train, labels_train)

print('输出分类模型的准确率...')
accuracy(model_svm, hog_vector_train, labels_train)

print('测试分类模型...')
result = image_predict(model_svm, data_test, hog_vector_test, labels_test)

cv2.imshow('result, press the q key to exit', result)
while 1:
if cv2.waitKey() == ord('q'):
break

print('测试其他图像...')
class_name = {0: "不包含行人", 1: "包含行人"}
img = cv2.imread('image\\persontrain.png')
img = cv2.resize(img, dsize=(64, 128))
img2 = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
img_preprocess = cv2.GaussianBlur(img2, (3, 3), sigmaX=1, sigmaY=1)
# vector = np.array([hog.compute(img_preprocess)])
vector = np.expand_dims(hog.compute(img_preprocess), 0)
ret = model_svm.predict(vector)[1].ravel()
print(f"图片img{class_name[int(ret)]}")

cv2.imshow('img', img)
cv2.waitKey()
cv2.destroyAllWindows()