问题

有一组水果的训练集，我们对模型进行训练，思路跟之前我们识别猫与狗一样。

设计解决这个问题的思路

1、下载与放置训练图片
2、现在对应的依赖，tensorflow、numpy等等
3、构建训练集合
4、建模
5、对模型进行训练
6、用测试模型进行验证
7、输出结果
8、优化模型 to step4

[1]图片地址

https://www.kaggle.com/moltean/fruits 现在数据，现在速度比较慢，可以使用网盘。

网盘地址(提取码:a9wr)

目录结构

【2】处理训练集的数据结构

import os
import pandas as pd

train_dir = './Training/'
test_dir = './Test/'
fruits = []
fruits_image = []

for i in os.listdir(train_dir):
    for image_filename in os.listdir(train_dir + i):
        fruits.append(i) # name of the fruit
        fruits_image.append(i + '/' + image_filename)
train_fruits = pd.DataFrame(fruits, columns=["Fruits"])
train_fruits["Fruits Image"] = fruits_image

print(train_fruits)

结果输出

           Fruits              Fruits Image
0        Tomato 4    Tomato 4/r_236_100.jpg
1        Tomato 4      Tomato 4/247_100.jpg
2        Tomato 4      Tomato 4/257_100.jpg
3        Tomato 4     Tomato 4/r_78_100.jpg
4        Tomato 4     Tomato 4/r_68_100.jpg
...           ...                       ...
67687  Peach Flat    Peach Flat/220_100.jpg
67688  Peach Flat  Peach Flat/r_127_100.jpg
67689  Peach Flat    Peach Flat/156_100.jpg
67690  Peach Flat  Peach Flat/r_137_100.jpg
67691  Peach Flat    Peach Flat/146_100.jpg

【3】构造模型

import matplotlib.pyplot as plt
import seaborn as sns
from keras.preprocessing.image import ImageDataGenerator, img_to_array, load_img
from glob import glob
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Activation, Dropout, Flatten, Dense
img = load_img(train_dir + "Cantaloupe 1/r_234_100.jpg")
plt.imshow(img)
plt.axis("off")
plt.show()

array_image = img_to_array(img)

# shape (100,100)
print("Image Shape --> ", array_image.shape)

# 131个类目
fruitCountUnique = glob(train_dir + '/*' )
numberOfClass = len(fruitCountUnique)
print("How many different fruits are there --> ",numberOfClass)

# 构建模型
model = Sequential()
model.add(Conv2D(32,(3,3),input_shape = array_image.shape))
model.add(Activation("relu"))
model.add(MaxPooling2D())
model.add(Conv2D(32,(3,3)))
model.add(Activation("relu"))
model.add(MaxPooling2D())
model.add(Conv2D(64,(3,3)))
model.add(Activation("relu"))
model.add(MaxPooling2D())
model.add(Flatten())
model.add(Dense(1024))
model.add(Activation("relu"))
model.add(Dropout(0.5))

# 区分131类
model.add(Dense(numberOfClass)) # output
model.add(Activation("softmax"))
model.compile(loss = "categorical_crossentropy",

              optimizer = "rmsprop",

              metrics = ["accuracy"])

print("Target Size --> ", array_image.shape[:2])

【4】训练模型

train_datagen = ImageDataGenerator(rescale= 1./255,
                                   shear_range = 0.3,
                                   horizontal_flip=True,
                                   zoom_range = 0.3)

test_datagen = ImageDataGenerator(rescale= 1./255)
epochs = 100
batch_size = 32
train_generator = train_datagen.flow_from_directory(
                train_dir,
                target_size= array_image.shape[:2],
                batch_size = batch_size,
                color_mode= "rgb",
                class_mode= "categorical")

test_generator = test_datagen.flow_from_directory(
                test_dir,
                target_size= array_image.shape[:2],
                batch_size = batch_size,
                color_mode= "rgb",
                class_mode= "categorical")

for data_batch, labels_batch in train_generator:
    print("data_batch shape --> ",data_batch.shape)
    print("labels_batch shape --> ",labels_batch.shape)
    break

hist = model.fit_generator(
        generator = train_generator,
        steps_per_epoch = 1600 // batch_size,
        epochs=epochs,
        validation_data = test_generator,
        validation_steps = 800 // batch_size)

#保存模型 model_fruits.h5
model.save('model_fruits.h5')

【5】展示训练结果

#展示损失模型结果
plt.figure()
plt.plot(hist.history["loss"],label = "Train Loss", color = "black")
plt.plot(hist.history["val_loss"],label = "Validation Loss", color = "darkred", linestyle="dashed",markeredgecolor = "purple", markeredgewidth = 2)
plt.title("Model Loss", color = "darkred", size = 13)
plt.legend()
plt.show()

#展示精确模型结果
plt.figure()
plt.plot(hist.history["accuracy"],label = "Train Accuracy", color = "black")
plt.plot(hist.history["val_accuracy"],label = "Validation Accuracy", color = "darkred", linestyle="dashed",markeredgecolor = "purple", markeredgewidth = 2)
plt.title("Model Accuracy", color = "darkred", size = 13)
plt.legend()
plt.show()

损失模型

验证模型

验证

新建一个文件，用于验证训练出来的模型

from tensorflow.keras.models import load_model
import os
import pandas as pd

from keras.preprocessing.image import ImageDataGenerator,img_to_array, load_img
import cv2,matplotlib.pyplot as plt,numpy as np
from keras.preprocessing import image

train_datagen = ImageDataGenerator(rescale= 1./255,
                                    shear_range = 0.3,
                                    horizontal_flip=True,
                                    zoom_range = 0.3)

model = load_model('model_fruits.h5')
batch_size = 32
img = load_img("./Test/Apricot/3_100.jpg",target_size=(100,100))
plt.imshow(img)
plt.show()

array_image = img_to_array(img)
array_image = array_image * 1./255
x = np.expand_dims(array_image, axis=0)
images = np.vstack([x])
classes = model.predict_classes(images, batch_size=10)
print(classes)
train_dir = './Training/'

train_generator = train_datagen.flow_from_directory(
        train_dir,
        target_size= array_image.shape[:2],
        batch_size = batch_size,
        color_mode= "rgb",
        class_mode= "categorical”)
print(train_generator.class_indices)

输出结果

[13]

Found 67692 images belonging to 131 classes.

{'Apple Braeburn': 0, 'Apple Crimson Snow': 1, 'Apple Golden 1': 2, 'Apple Golden 2': 3, 'Apple Golden 3': 4, 'Apple Granny Smith': 5, 'Apple Pink Lady': 6, 'Apple Red 1': 7, 'Apple Red 2': 8, 'Apple Red 3': 9, 'Apple Red Delicious': 10, 'Apple Red Yellow 1': 11, 'Apple Red Yellow 2': 12, 'Apricot': 13, 'Avocado': 14, 'Avocado ripe': 15, 'Banana': 16, 'Banana Lady Finger': 17, 'Banana Red': 18, 'Beetroot': 19, 'Blueberry': 20, 'Cactus fruit': 21, 'Cantaloupe 1': 22, 'Cantaloupe 2': 23, 'Carambula': 24, 'Cauliflower': 25, 'Cherry 1': 26, 'Cherry 2': 27, 'Cherry Rainier': 28, 'Cherry Wax Black': 29, 'Cherry Wax Red': 30, 'Cherry Wax Yellow': 31, 'Chestnut': 32, 'Clementine': 33, 'Cocos': 34, 'Corn': 35, 'Corn Husk': 36, 'Cucumber Ripe': 37, 'Cucumber Ripe 2': 38, 'Dates': 39, 'Eggplant': 40, 'Fig': 41, 'Ginger Root': 42, 'Granadilla': 43, 'Grape Blue': 44, 'Grape Pink': 45, 'Grape White': 46, 'Grape White 2': 47, 'Grape White 3': 48, 'Grape White 4': 49, 'Grapefruit Pink': 50, 'Grapefruit White': 51, 'Guava': 52, 'Hazelnut': 53, 'Huckleberry': 54, 'Kaki': 55, 'Kiwi': 56, 'Kohlrabi': 57, 'Kumquats': 58, 'Lemon': 59, 'Lemon Meyer': 60, 'Limes': 61, 'Lychee': 62, 'Mandarine': 63, 'Mango': 64, 'Mango Red': 65, 'Mangostan': 66, 'Maracuja': 67, 'Melon Piel de Sapo': 68, 'Mulberry': 69, 'Nectarine': 70, 'Nectarine Flat': 71, 'Nut Forest': 72, 'Nut Pecan': 73, 'Onion Red': 74, 'Onion Red Peeled': 75, 'Onion White': 76, 'Orange': 77, 'Papaya': 78, 'Passion Fruit': 79, 'Peach': 80, 'Peach 2': 81, 'Peach Flat': 82, 'Pear': 83, 'Pear 2': 84, 'Pear Abate': 85, 'Pear Forelle': 86, 'Pear Kaiser': 87, 'Pear Monster': 88, 'Pear Red': 89, 'Pear Stone': 90, 'Pear Williams': 91, 'Pepino': 92, 'Pepper Green': 93, 'Pepper Orange': 94, 'Pepper Red': 95, 'Pepper Yellow': 96, 'Physalis': 97, 'Physalis with Husk': 98, 'Pineapple': 99, 'Pineapple Mini': 100, 'Pitahaya Red': 101, 'Plum': 102, 'Plum 2': 103, 'Plum 3': 104, 'Pomegranate': 105, 'Pomelo Sweetie': 106, 'Potato Red': 107, 'Potato Red Washed': 108, 'Potato Sweet': 109, 'Potato White': 110, 'Quince': 111, 'Rambutan': 112, 'Raspberry': 113, 'Redcurrant': 114, 'Salak': 115, 'Strawberry': 116, 'Strawberry Wedge': 117, 'Tamarillo': 118, 'Tangelo': 119, 'Tomato 1': 120, 'Tomato 2': 121, 'Tomato 3': 122, 'Tomato 4': 123, 'Tomato Cherry Red': 124, 'Tomato Heart': 125, 'Tomato Maroon': 126, 'Tomato Yellow': 127, 'Tomato not Ripened': 128, 'Walnut': 129, 'Watermelon': 130}

识别出是13：Apricot

进一步验证

fig = plt.figure(figsize=(16, 16))
axes = []
files = []
predictions = []
true_labels = []
rows = 5
cols = 2

# 随机选择几个图片
def getRandomImage(path, img_width, img_height):
    """function loads a random image from a random folder in our test path"""
    folders = list(filter(lambda x: os.path.isdir(os.path.join(path, x)), os.listdir(path)))
    random_directory = np.random.randint(0, len(folders))
    path_class = folders[random_directory]
    file_path = os.path.join(path, path_class)
    file_names = [f for f in os.listdir(file_path) if os.path.isfile(os.path.join(file_path, f))]
    random_file_index = np.random.randint(0, len(file_names))
    image_name = file_names[random_file_index]
    final_path = os.path.join(file_path, image_name)
    return image.load_img(final_path, target_size = (img_width, img_height)), final_path, path_class

def draw_test(name, pred, im, true_label):
    BLACK = [0, 0, 0]
    expanded_image = cv2.copyMakeBorder(im, 160, 0, 0, 300, cv2.BORDER_CONSTANT, value=BLACK)
    cv2.putText(expanded_image, "predicted: " + pred, (20, 60), cv2.FONT_HERSHEY_SIMPLEX,
        0.85, (255, 0, 0), 2)
    cv2.putText(expanded_image, "true: " + true_label, (20, 120), cv2.FONT_HERSHEY_SIMPLEX,
        0.85, (0, 255, 0), 2)
    return expanded_image
IMG_ROWS, IMG_COLS = 100, 100

# predicting images
for i in range(0, 10):
    path = "./Test"
    img, final_path, true_label = getRandomImage(path, IMG_ROWS, IMG_COLS)
    files.append(final_path)
    true_labels.append(true_label)
    x = image.img_to_array(img)
    x = x * 1./255
    x = np.expand_dims(x, axis=0)
    images = np.vstack([x])
    classes = model.predict_classes(images, batch_size=10)
    predictions.append(classes)

class_labels = train_generator.class_indices
class_labels = {v: k for k, v in class_labels.items()}
class_list = list(class_labels.values())

for i in range(0, len(files)):
    image = cv2.imread(files[i])
    image = draw_test("Prediction", class_labels[predictions[i][0]], image, true_labels[i])
    axes.append(fig.add_subplot(rows, cols, i+1))
    plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
    plt.grid(False)
    plt.axis('off')
plt.show()

呈现结果

水果识别