基于深度学习的交通标识别算法对比研究-TensorFlow2实现

• 🔗 运行环境：python3
• 🚩 作者：K同学啊
• 🥇 精选专栏：《深度学习100例》
• 🔥 推荐专栏：《零基础入门深度学习》
• 📚 选自专栏：《Matplotlib教程》
• 🧿 优秀专栏：《Python入门100题》

大家好，我是K同学啊！

今天和大家分享一篇 本科毕设 实战项目，项目中我将使用VGG16、InceptionV3、DenseNet121、MobileNetV2 等四个模型进行对比分析（文中提供了每一个模型的 算法框架图），最后可以自选图片进行预测，最后的识别效果高达 99.2%。结果如下：

@[toc]

一、导入数据

""" 关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/117018789 """
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "./1-data/",
    validation_split=0.2,
    subset="training",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 1308 files belonging to 14 classes.
Using 1047 files for training.

""" 关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/117018789 """
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "./1-data/",
    validation_split=0.2,
    subset="validation",
    seed=12,
    image_size=(img_height, img_width),
    batch_size=batch_size)

Found 1308 files belonging to 14 classes.
Using 261 files for validation.

class_names = train_ds.class_names
print(class_names)

['15', '16', '17', '20', '22', '23', '24', '26', '27', '28', '29', '30', '31', '32']

train_ds

<BatchDataset shapes: ((None, 224, 224, 3), (None,)), types: (tf.float32, tf.int32)>

AUTOTUNE = tf.data.AUTOTUNE

# 归一化
def train_preprocessing(image,label):
    return (image/255.0,label)

train_ds = (
    train_ds.cache()
    .map(train_preprocessing)    # 这里可以设置预处理函数
    .prefetch(buffer_size=AUTOTUNE)
)

val_ds = (
    val_ds.cache()
    .map(train_preprocessing)    # 这里可以设置预处理函数
    .prefetch(buffer_size=AUTOTUNE)
)

plt.figure(figsize=(10, 8))  # 图形的宽为10高为5

for images, labels in train_ds.take(1):
    for i in range(15):
        plt.subplot(4, 5, i + 1)
        plt.xticks([])
        plt.yticks([])
        plt.grid(False)

        # 显示图片
        plt.imshow(images[i])
        # 显示标签
        plt.xlabel(class_names[int(labels[i])])

plt.show()

二、定义模型

1. VGG16模型

# 加载预训练模型
vgg16_base_model = tf.keras.applications.vgg16.VGG16(weights='imagenet',
                                                            include_top=False,
                                                            # input_tensor=tf.keras.Input(shape=(img_width, img_height, 3)),
                                                            input_shape=(img_width, img_height, 3),
                                                            pooling='max')
for layer in vgg16_base_model.layers:
    layer.trainable = False
    
X = vgg16_base_model.output
X = Dropout(0.4)(X)

output = Dense(len(class_names), activation='softmax')(X)
vgg16_model = Model(inputs=vgg16_base_model.input, outputs=output)

vgg16_model.compile(optimizer="adam",
                loss='sparse_categorical_crossentropy',
                metrics=['accuracy'])
# vgg16_model.summary()

vgg16_history  = vgg16_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)

Epoch 1/10
33/33 [==============================] - 8s 113ms/step - loss: 2.7396 - accuracy: 0.2531 - val_loss: 1.4678 - val_accuracy: 0.6092
Epoch 2/10
33/33 [==============================] - 2s 45ms/step - loss: 1.5873 - accuracy: 0.5091 - val_loss: 0.8500 - val_accuracy: 0.8046
Epoch 3/10
33/33 [==============================] - 2s 45ms/step - loss: 1.0996 - accuracy: 0.6495 - val_loss: 0.5299 - val_accuracy: 0.9272
Epoch 4/10
33/33 [==============================] - 2s 45ms/step - loss: 0.7349 - accuracy: 0.7947 - val_loss: 0.3765 - val_accuracy: 0.9349
Epoch 5/10
33/33 [==============================] - 2s 45ms/step - loss: 0.5373 - accuracy: 0.8481 - val_loss: 0.2888 - val_accuracy: 0.9502
Epoch 6/10
33/33 [==============================] - 2s 45ms/step - loss: 0.4326 - accuracy: 0.8892 - val_loss: 0.2422 - val_accuracy: 0.9617
Epoch 7/10
33/33 [==============================] - 2s 45ms/step - loss: 0.3350 - accuracy: 0.9198 - val_loss: 0.2068 - val_accuracy: 0.9693
Epoch 8/10
33/33 [==============================] - 2s 45ms/step - loss: 0.2821 - accuracy: 0.9398 - val_loss: 0.1713 - val_accuracy: 0.9885
Epoch 9/10
33/33 [==============================] - 2s 45ms/step - loss: 0.2489 - accuracy: 0.9456 - val_loss: 0.1589 - val_accuracy: 0.9847
Epoch 10/10
33/33 [==============================] - 2s 48ms/step - loss: 0.2146 - accuracy: 0.9608 - val_loss: 0.1511 - val_accuracy: 0.9885

2. InceptionV3模型

# 加载预训练模型
InceptionV3_base_model = tf.keras.applications.inception_v3.InceptionV3(weights='imagenet',
                                                            include_top=False,
                                                            input_shape=(img_width, img_height, 3),
                                                            pooling='max')
for layer in InceptionV3_base_model.layers:
    layer.trainable = False
    
X = InceptionV3_base_model.output
X = Dropout(0.4)(X)

output = Dense(len(class_names), activation='softmax')(X)
InceptionV3_model = Model(inputs=InceptionV3_base_model.input, outputs=output)

InceptionV3_model.compile(optimizer="adam",
                loss='sparse_categorical_crossentropy',
                metrics=['accuracy'])
# InceptionV3_model.summary()

InceptionV3_history  = InceptionV3_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)

Epoch 1/10
33/33 [==============================] - 5s 82ms/step - loss: 3.1642 - accuracy: 0.4040 - val_loss: 0.6005 - val_accuracy: 0.8352
Epoch 2/10
33/33 [==============================] - 1s 34ms/step - loss: 0.7241 - accuracy: 0.8042 - val_loss: 0.2476 - val_accuracy: 0.9234
Epoch 3/10
33/33 [==============================] - 1s 34ms/step - loss: 0.3558 - accuracy: 0.8949 - val_loss: 0.2323 - val_accuracy: 0.9425
Epoch 4/10
33/33 [==============================] - 1s 35ms/step - loss: 0.2435 - accuracy: 0.9226 - val_loss: 0.1599 - val_accuracy: 0.9617
Epoch 5/10
33/33 [==============================] - 1s 34ms/step - loss: 0.1444 - accuracy: 0.9551 - val_loss: 0.1246 - val_accuracy: 0.9617
Epoch 6/10
33/33 [==============================] - 1s 34ms/step - loss: 0.1508 - accuracy: 0.9522 - val_loss: 0.1231 - val_accuracy: 0.9732
Epoch 7/10
33/33 [==============================] - 1s 35ms/step - loss: 0.0793 - accuracy: 0.9761 - val_loss: 0.0853 - val_accuracy: 0.9885
Epoch 8/10
33/33 [==============================] - 1s 35ms/step - loss: 0.0636 - accuracy: 0.9809 - val_loss: 0.1223 - val_accuracy: 0.9732
Epoch 9/10
33/33 [==============================] - 1s 35ms/step - loss: 0.0503 - accuracy: 0.9857 - val_loss: 0.0769 - val_accuracy: 0.9923
Epoch 10/10
33/33 [==============================] - 1s 34ms/step - loss: 0.0346 - accuracy: 0.9904 - val_loss: 0.1066 - val_accuracy: 0.9923

3. DenseNet121算法模型

# 加载预训练模型
DenseNet121_base_model = tf.keras.applications.densenet.DenseNet121(weights='imagenet',
                                                            include_top=False,
                                                            input_shape=(img_width, img_height, 3),
                                                            pooling='max')
for layer in DenseNet121_base_model.layers:
    layer.trainable = False
    
X = DenseNet121_base_model.output
X = Dropout(0.4)(X)

output = Dense(len(class_names), activation='softmax')(X)
DenseNet121_model = Model(inputs=DenseNet121_base_model.input, outputs=output)

DenseNet121_model.compile(optimizer="adam",
                loss='sparse_categorical_crossentropy',
                metrics=['accuracy'])
# model.summary()

DenseNet121_history  = DenseNet121_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)

Epoch 1/10
33/33 [==============================] - 7s 109ms/step - loss: 4.5573 - accuracy: 0.2932 - val_loss: 1.2358 - val_accuracy: 0.6322
Epoch 2/10
33/33 [==============================] - 1s 43ms/step - loss: 2.0711 - accuracy: 0.5482 - val_loss: 0.4970 - val_accuracy: 0.8391
Epoch 3/10
33/33 [==============================] - 1s 41ms/step - loss: 1.2808 - accuracy: 0.6953 - val_loss: 0.2534 - val_accuracy: 0.9042
Epoch 4/10
33/33 [==============================] - 1s 41ms/step - loss: 0.8280 - accuracy: 0.7736 - val_loss: 0.1845 - val_accuracy: 0.9502
Epoch 5/10
33/33 [==============================] - 1s 41ms/step - loss: 0.5928 - accuracy: 0.8300 - val_loss: 0.1211 - val_accuracy: 0.9770
Epoch 6/10
33/33 [==============================] - 1s 41ms/step - loss: 0.4390 - accuracy: 0.8749 - val_loss: 0.1046 - val_accuracy: 0.9808
Epoch 7/10
33/33 [==============================] - 1s 41ms/step - loss: 0.4108 - accuracy: 0.8797 - val_loss: 0.0950 - val_accuracy: 0.9885
Epoch 8/10
33/33 [==============================] - 1s 41ms/step - loss: 0.3137 - accuracy: 0.9102 - val_loss: 0.0662 - val_accuracy: 0.9808
Epoch 9/10
33/33 [==============================] - 1s 41ms/step - loss: 0.2416 - accuracy: 0.9284 - val_loss: 0.0698 - val_accuracy: 0.9885
Epoch 10/10
33/33 [==============================] - 1s 41ms/step - loss: 0.2524 - accuracy: 0.9217 - val_loss: 0.0597 - val_accuracy: 0.9923

4. MobileNetV2算法模型

# 加载预训练模型
MobileNetV2_base_model = tf.keras.applications.mobilenet_v2.MobileNetV2(weights='imagenet',
                                                            include_top=False,
                                                            input_shape=(img_width, img_height, 3),
                                                            pooling='max')
for layer in MobileNetV2_base_model.layers:
    layer.trainable = False
    
X = MobileNetV2_base_model.output
X = Dropout(0.4)(X)

output = Dense(len(class_names), activation='softmax')(X)
MobileNetV2_model = Model(inputs=MobileNetV2_base_model.input, outputs=output)

MobileNetV2_model.compile(optimizer="adam",
                loss='sparse_categorical_crossentropy',
                metrics=['accuracy'])
#MobileNetV2_model.summary()

MobileNetV2_history  = MobileNetV2_model.fit(train_ds, epochs=epochs, verbose=1, validation_data=val_ds)

Epoch 1/10
33/33 [==============================] - 3s 47ms/step - loss: 4.0865 - accuracy: 0.4403 - val_loss: 0.5897 - val_accuracy: 0.8812
Epoch 2/10
33/33 [==============================] - 1s 22ms/step - loss: 1.1042 - accuracy: 0.7536 - val_loss: 0.1841 - val_accuracy: 0.9540
Epoch 3/10
33/33 [==============================] - 1s 22ms/step - loss: 0.6147 - accuracy: 0.8596 - val_loss: 0.1722 - val_accuracy: 0.9770
Epoch 4/10
33/33 [==============================] - 1s 22ms/step - loss: 0.3826 - accuracy: 0.9007 - val_loss: 0.1505 - val_accuracy: 0.9770
Epoch 5/10
33/33 [==============================] - 1s 22ms/step - loss: 0.2290 - accuracy: 0.9370 - val_loss: 0.1408 - val_accuracy: 0.9885
Epoch 6/10
33/33 [==============================] - 1s 22ms/step - loss: 0.1976 - accuracy: 0.9484 - val_loss: 0.1294 - val_accuracy: 0.9923
Epoch 7/10
33/33 [==============================] - 1s 22ms/step - loss: 0.1193 - accuracy: 0.9608 - val_loss: 0.1038 - val_accuracy: 0.9923
Epoch 8/10
33/33 [==============================] - 1s 22ms/step - loss: 0.0859 - accuracy: 0.9675 - val_loss: 0.1140 - val_accuracy: 0.9923
Epoch 9/10
33/33 [==============================] - 1s 22ms/step - loss: 0.0973 - accuracy: 0.9704 - val_loss: 0.1292 - val_accuracy: 0.9923
Epoch 10/10
33/33 [==============================] - 1s 22ms/step - loss: 0.0504 - accuracy: 0.9828 - val_loss: 0.1361 - val_accuracy: 0.9923

三、结果分析

1. 准确率对比分析

# 可在原码中进行阅读
plt.show()

2. 损失函数对比分析

# 可在原码中进行阅读
plt.show()

3. 混淆矩阵

# 可在原码中进行阅读
plot_cm(val_label, val_pre)

4. 评估指标生成

• support：当前行的类别在测试数据中的样本总量；
• precision：被判定为正例（反例）的样本中，真正的正例样本（反例样本）的比例，精度=正确预测的个数(TP)/被预测正确的个数(TP+FP)。
• recall：被正确分类的正例（反例）样本，占所有正例（反例）样本的比例，召回率=正确预测的个数(TP)/预测个数(TP+FN)。
• f1-score: 精确率和召回率的调和平均值，F1 = 2精度召回率/(精度+召回率)。
• accuracy：表示准确率，也即正确预测样本量与总样本量的比值。
• macro avg：表示宏平均，表示所有类别对应指标的平均值。
• weighted avg：表示带权重平均，表示类别样本占总样本的比重与对应指标的乘积的累加和。

from sklearn import metrics

def test_accuracy_report(model):
    print(metrics.classification_report(val_label, val_pre, target_names=class_names)) 
    score = model.evaluate(val_ds, verbose=0)
    print('Loss function: %s, accuracy:' % score[0], score[1])

test_accuracy_report(InceptionV3_model)

 precision recall f1-score support 
 15 1.00 1.00 1.00 2 16 1.00 1.00 1.00 28 17 1.00 1.00 1.00 25 20 1.00 0.33 0.50 3 22 1.00 1.00 1.00 4 23 1.00 1.00 1.00 1 24 1.00 1.00 1.00 16 26 1.00 1.00 1.00 32 27 0.71 1.00 0.83 5 28 1.00 1.00 1.00 90 29 1.00 1.00 1.00 5 30 1.00 1.00 1.00 33 31 1.00 1.00 1.00 8 32 1.00 1.00 1.00 9 
 accuracy 0.99 261 macro avg 0.98 0.95 0.95 261 weighted avg 0.99 0.99 0.99 261 
Loss function: 0.10659126937389374, accuracy: 0.992337167263031

四、指定图片进行预测

from PIL import Image

img = Image.open("./1-data/17/017_0001.png")
image = tf.image.resize(img, [img_height, img_width])

img_array = tf.expand_dims(image, 0) 

predictions = InceptionV3_model.predict(img_array)
print("预测结果为：",np.argmax(predictions))

预测结果为： 11

完整资源下载链接1：博主在面包多网站上的完整资源下载页

完整资源下载链接1：mianbaoduo.com/o/bread/Ypi…

🎁 粉丝福利

书籍系列

• 🎈 机器学习-周志华 百度云下载链接 提取码：hnzu
• 🎆 统计学习-李航 百度云下载链接 提取码：lc5e
• 🧨 神经网络与深度学习-邱锡鹏 百度云下载链接 提取码：m3ls
• ✨ 深度强化学习 Hung-yi Lee 百度云下载链接 提取码：9b0a
• 🎉 机器学习实战 中文双页版 百度云下载链接 提取码：jy7r
• 🎊 斯坦福大学-深度学习基础教程 百度云下载链接 提取码：hjz7
• 🎃 Tensorflow 实战Google深度学习框架百度云下载链接 提取码：f6wf

面试笔记系列

• 🎄 谷歌和阿里大佬的Leetcode刷题笔记 百度云下载链接 提取码：9go2
• 🎋 Leetcode 1300道算法题解 百度云下载链接 提取码：eg9z
• 🎍 labuladong 的算法小抄 百度云下载链接 提取码：v4qf
• 🎎 程序员面试宝典 百度云下载链接 提取码：gm7o
• 🎏 大厂面试真题 百度云下载链接 提取码：u9y4