First of all, I prepared five categories of pictures in advance and put them in the corresponding folders. The picture resources are linked in my gitee folder as follows: File management: used to store various data https://gitee.com/xiaoxiaotai/file-management .git
There are imgs directory and npy directory inside. imgs is the directory that stores 5 categories of pictures. There are 5 kinds of plants in it, osmanthus, maple leaf, schisandra, ginkgo, and bamboo leaves. The npy directory stores the npy file data set I made with these pictures. , which has npy files of 32x32 size and 64x64 size.
Next is the dataset creation process:
First import the required libraries
import os
import cv2
import random
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
from mpl_toolkits.axes_grid1 import ImageGrid
%matplotlib inline
import math
from tqdm import tqdm
The following is the first display of some pictures in the local classification
#先显示枫叶图片
folder_path = './datas/imgs/fengye'
# 可视化图像的个数
N = 36
# n 行 n 列
n = math.floor(np.sqrt(N))
images = []
for each_img in os.listdir(folder_path)[:N]:
img_path = os.path.join(folder_path, each_img)
#img_bgr = cv2.imread(img_path)
img_bgr = cv2.imdecode(np.fromfile(img_path, dtype=np.uint8), 1) #解决路径中存在中文的问题
img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
images.append(img_rgb)
fig = plt.figure(figsize=(6, 8),dpi=80)
grid = ImageGrid(fig, 111, # 类似绘制子图 subplot(111)
nrows_ncols=(n, n), # 创建 n 行 m 列的 axes 网格
axes_pad=0.02, # 网格间距
share_all=True
)
# 遍历每张图像
for ax, im in zip(grid, images):
ax.imshow(im)
ax.axis('off')
plt.tight_layout()
plt.show()
The output is as follows:
The following is to output the information of each picture, including picture width and height, picture name, category, os.chdir('../') means to point the current path pointer to the previous directory, you can use os.getcwd() to output the current point path
# 指定数据集路径
dataset_path = './datas/imgs/'
os.chdir(dataset_path)
print(os.listdir())
df = pd.DataFrame()
for fruit in tqdm(os.listdir()): # 遍历每个类别
os.chdir(fruit)
for file in os.listdir(): # 遍历每张图像
try:
img = cv2.imread(file)
df = df.append({'类别':fruit, '文件名':file, '图像宽':img.shape[1], '图像高':img.shape[0]}, ignore_index=True)
except:
print(os.path.join(fruit, file), '读取错误')
os.chdir('../')
os.chdir('../../')
dfThe output is as follows:
Define the label number, because the data set label is generally a number, the training is faster
# 定义5个类别的标签
labels = {
'wuweizi': 0,
'fengye': 1,
'guihua': 2,
'zhuye': 3,
'yinxing': 4
}
# 定义训练集和测试集的比例
train_ratio = 0.8
# 定义一个空列表用于存储训练集和测试集
train_data = []
test_data = []For data enhancement, here I am reducing each picture to 64x64. You can also change it to 32x32 or other sizes. It should be noted that the larger the size, the longer the data set is produced, and the larger the size of the obtained data set.
# 定义数据增强的方法
def data_augmentation(img):
# 随机裁剪
img = cv2.resize(img, (256, 256))
x = random.randint(0, 256 - 64)
y = random.randint(0, 256 - 64)
img = img[x:x+64, y:y+64]
# 随机翻转
if random.random() < 0.5:
img = cv2.flip(img, 1)
# 随机旋转
angle = random.randint(-10, 10)
M = cv2.getRotationMatrix2D((32, 32), angle, 1)
img = cv2.warpAffine(img, M, (64, 64))
return img# 定义读取图片的方法
def read_image(path):
img = cv2.imread(path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = data_augmentation(img)
img = img / 255.0
return imgThe following is to label the pictures, that is, each picture is marked with which category (identity) it belongs to, so that the convolutional neural network can know the category during training, so as to remember the label value of the feature
# 遍历5个文件夹，读取图片并打上标签
for path, label in labels.items():
files = os.listdir('./datas/imgs/'+path)
random.shuffle(files)
train_files = files[:int(len(files) * train_ratio)]
test_files = files[int(len(files) * train_ratio):]
for file in train_files:
img = read_image(os.path.join('./datas/imgs/'+path, file))
train_data.append((img, label))
for file in test_files:
img = read_image(os.path.join('./datas/imgs/'+path, file))
test_data.append((img, label))
# 工整地输出每一类别的数据个数
print('类别：{} 训练集个数：{} 测试集数据：{}'.format(path, len(train_files), len(test_files)))Here's the output:
Now you can look at the cropped result
df = pd.DataFrame()
for img,label in train_data: # 遍历每个类别
# img = cv2.imread(fruit)
df = df.append({'类别':label, '文件名':file, '图像宽':img.shape[1], '图像高':img.shape[0]}, ignore_index=True)
dfThe result is as follows, we can see that the size has become 64x64, of course, this is not in order, the category is from 0 to 4:
The next step is to shuffle the order, which is also to prevent overfitting
# 打乱训练集和测试集的顺序
random.shuffle(train_data)
random.shuffle(test_data)output again
df = pd.DataFrame()
for img,label in train_data: # 遍历每个类别
# img = cv2.imread(fruit)
df = df.append({'类别':label, '文件名':file, '图像宽':img.shape[1], '图像高':img.shape[0]}, ignore_index=True)
dfThis time the results are as follows, the order of categories has been shuffled:
The following are the datasets and labels that hold the training and test sets
# 将训练集和测试集的图片和标签分别存储在numpy数组中
train_imgs = np.array([data[0] for data in train_data])
train_labels = np.array([data[1] for data in train_data])
test_imgs = np.array([data[0] for data in test_data])
test_labels = np.array([data[1] for data in test_data])
# 保存训练集和测试集
np.save('./datas/npy/32px/train_imgs_64.npy', train_imgs)
np.save('./datas/npy/32px/train_labels_64.npy', train_labels)
np.save('./datas/npy/32px/test_imgs_64.npy', test_imgs)
np.save('./datas/npy/32px/test_labels_64.npy', test_labels)The above data set has been prepared, so the next step is the training of the model. I will not explain the training of the model one by one. You can look at the code yourself. I use the jupyter tool in anaconda to write the code
#导库
import tensorflow as tf
import numpy as np
import os
import matplotlib.pyplot as plt
import urllib
import cv2
# 加载上面制作的数据集
train_imgs = np.load('./datas/npy/64px/train_imgs_64.npy')
train_labels = np.load('./datas/npy/64px/train_labels_64.npy')
test_imgs = np.load('./datas/npy/64px/test_imgs_64.npy')
test_labels = np.load('./datas/npy/64px/test_labels_64.npy')
#可以看看输出纬度
train_imgs.shape
#模型构建，这里我就构建一个简单模型
def creatAlexNet():
model = tf.keras.models.Sequential([
tf.keras.layers.Conv2D(64, kernel_size=(3, 3), strides=(1, 1), activation='relu', input_shape=(64, 64, 3)),
tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(1, 1)),
tf.keras.layers.Conv2D(128, kernel_size=(3, 3), strides=(1, 1), activation='relu'),
tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(1, 1)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(5, activation='softmax')
])
return model
#加载模型
model = creatAlexNet()
#显示摘要
model.summary()
# 定义超参数
learning_rate = 0.001 #study
batch_size = 100 #单次训练样本数（批次大小）
epochs = 20 #训练轮数
# 定义训练模式
model.compile(optimizer ='adam',#优化器
loss='sparse_categorical_crossentropy',#损失函数
metrics=['accuracy'])#评估模型的方式
# 加载数据集并训练模型
history = model.fit(train_imgs, train_labels, batch_size=batch_size, epochs=epochs,
validation_split = 0.2)
# 评估模型
test_loss, test_acc = model.evaluate(test_imgs, test_labels, verbose=2)
print('Test accuracy:', test_acc)
#模型测试
preds = model.predict(test_imgs)
np.argmax(preds[20])
# 可视化测试
# 定义显示图像数据及其对应标签的函数
# 图像列表
label_dict={0:"wuweizi",1:"fengye",2:"guihua",3:"zhuye",4:"yinxing"}
def plot_images_labels_prediction(images,# 标签列表
labels,
preds,#预测值列表
index,#从第index个开始显示
num = 5): # 缺省一次显示5幅
fig=plt.gcf()#获取当前图表，Get Current Figure
fig.set_size_inches(12,6)#1英寸等于2.54cm
if num > 10:#最多显示10个子图
num = 10
for i in range(0, num):
ax = plt.subplot(2,5,i+1)#获取当前要处理的子图
plt.tight_layout()
ax.imshow(images[index])
title=str(i)+','+label_dict[labels[index]]#构建该图上要显示的title信息
if len(preds)>0:
title +='=>' + label_dict[np.argmax(preds[index])]
ax.set_title(title,fontsize=10)#显示图上的title信息
index += 1
plt.show()
plot_images_labels_prediction(test_imgs,test_labels, preds,10,30)
# 然后保存模型
model_filename ='models/plant_model.h5'
model.save(model_filename)
# 这里是从本地加载图片对模型进行测试
from PIL import Image
import numpy as np
loaded_model = tf.keras.models.load_model('models/plant_model.h5')
label_dict={0:"wuweizi",1:"fengye",2:"guihua",3:"zhuye",4:"yinxing"}
img = Image.open('./fengye.jpeg')
img = img.resize((64, 64))
img_arr = np.array(img) / 255.0
img_arr = img_arr.reshape(1, 64, 64, 3)
pred = model.predict(img_arr)
class_idx = np.argmax(pred)
plt.title("type:{}, pre_label:{}".format(label_dict[class_idx],class_idx))
plt.imshow(img, cmap=plt.get_cmap('gray'))
# 加载模型
loaded_model = tf.keras.models.load_model('models/plant_model.h5')
# 使用模型预测浏览器上的一张图片
label_dict={0:"wuweizi",1:"fengye",2:"guihua",3:"zhuye",4:"yinxing"}
# 这里是从浏览器的网址中加载图片进行识别
url = 'https://newbbs-fd.zol-img.com.cn/t_s1200x5000/g5/M00/05/08/ChMkJ1wFsOGIcMt4AAGFQDPiUhEAAtkTQCj_EoAAYVY306.jpg'
with urllib.request.urlopen(url) as url_response:
img_array = np.asarray(bytearray(url_response.read()), dtype=np.uint8)
img = cv2.imdecode(img_array, cv2.IMREAD_COLOR)
img_array = cv2.resize(img, (64, 64))
img_array = img_array / 255.0
img_array = np.expand_dims(img_array, axis=0)
predict_label = np.argmax(loaded_model.predict(img_array), axis=-1)[0]
plt.imshow(img, cmap=plt.get_cmap('gray'))
plt.title("Predict: {},Predict_label: {}".format(label_dict[predict_label],predict_label))
plt.xticks([])
plt.yticks([])
This article is here, thank you for your support!