机器学习笔记3 交通标志识别 CNN with Keras

这篇文章中，我们会利用简单的CNN来训练出一个可以识别交通标志牌的模型。

首先，我们需要有一些数据集合， 下载链接如下：

datasets:

training:

https://btsd.ethz.ch/shareddata/BelgiumTSC/BelgiumTSC_Training.zip

testing:

https://btsd.ethz.ch/shareddata/BelgiumTSC/BelgiumTSC_Testing.zip

这些数据都有一点年头了，下载之后我发现这些数据集都是以ppm格式保存的，那么我们的第一步工作就是把图片转换成jpg，幸运的是，python中的PIL包集成了这个功能，代码如下：

def ppm2jpg(original_name, target_name):
    img = Image.open(original_name)
    img = img.resize((64, 64), Image.ANTIALIAS)
    img.save(target_name)

这个文件夹中的图片尺寸不一，我们统一放缩成 ‘’‘64*64’’'的图片。original_name指的是原始图片位置， 而target_name是要另存为的位置。
这里可能还需要创建文件夹， 代码如下 （这段代码可选， 因为我们也可以手动创建文件夹，毕竟就只有两个）

def mkdir(path):
    path=path.strip()
    path=path.rstrip("\\")
    isExists=os.path.exists(path)
    if not isExists:
        os.makedirs(path)
    else:
        return False

当准备好了这两个函数之后，我们便可以开始进行转换。
设置路径：

Training_folder = './BelgiumTSC_Training/Training'
Testing_folder = './BelgiumTSC_Testing/Testing'
Training_folder_new = './Training_jpg'
Testing_folder_new = './Testing_jpg'

直接放在当前目录下是最方便的。
批量转换函数：

    class_num = 0
    train_dir = os.listdir(original_folder)
    for item in train_dir:
        pic_folders = os.path.join(original_folder,item)
        pic_names = os.listdir(pic_folders)
        count = 0
        for i in range(len(pic_names) - 1):
            original_name = os.path.join(pic_folders, pic_names[i])
            target_dir = os.path.join(target_folder,str(class_num))
            mkdir(target_dir)
            target_name = os.path.join(target_dir, str(count) + ".jpg")
            ppm2jpg(original_name = original_name, target_name = target_name)
            count += 1
        class_num += 1

这里我还顺便分了类别，一开始的文件名并不是很友好，我这里直接设置成了类别。 类别0 的图片放在文件夹0 里面。
最后我们直接调用函数：

convert(original_folder = Training_folder, target_folder = Training_folder_new)
convert(original_folder = Testing_folder, target_folder = Testing_folder_new)`

---

这样子数据就处理完毕了。
在开始训练之前，我们需要将jpg图片读取出来并转换成array。同时还需要进行one hot转换， 这些都不是很难。

folder = ['./Training_jpg', './Testing_jpg']
def jpg_to_array(name):
    if name == 'train':
        idx = 0
    else:
        idx = 1
    feature = []
    label = []
    class_num = 62

    dir = os.listdir(folder[idx])
    dir.sort(key=lambda x: int(x[:]))
    for item in dir:
        hot = int(item)
        y = np.zeros(class_num)
        y[hot] = 1

        jpg_path = os.path.join(folder[idx], item)
        pic_list = os.listdir(jpg_path)
        for pic in pic_list:
            x = Image.open(os.path.join(jpg_path, pic))
            feature.append(np.array(x))
            label.append(y)
    features = np.array(feature)
    labels = np.array(label)
    return features, labels

调用函数：

features_train, labels_train = jpg_to_array('train')
features_test, labels_test = jpg_to_array('test')
print(features_train.shape)
print(features_test.shape)
print(labels_train.shape)
print(labels_test.shape)

我们得到如下结果：

(4575, 64, 64, 3)
(2520, 64, 64, 3)
(4575, 62)
(2520, 62)

---

然后开始设计简单的CNN 模型。 keras模块就和搭积木一样，很方便。 这里的模型其实挺随意的，个人觉得是因为训练集太简单了。

dropout_rate = 0.3
batch_size = 256
epochs = 20

---

model = keras.Sequential()

model.add(Conv2D(16, (3, 3), padding='same', input_shape=features_train.shape[1:], activation = 'relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(dropout_rate))
model.add(BatchNormalization())


model.add(Conv2D(32, (3, 3), padding='same', activation = 'relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(dropout_rate))
model.add(BatchNormalization())

model.add(Conv2D(64, (3, 3), padding='same', activation = 'relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(dropout_rate))
model.add(BatchNormalization())

model.add(Conv2D(128, (3, 3), padding='same', activation = 'relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(dropout_rate))
model.add(BatchNormalization())

model.add(Conv2D(256, (3, 3), padding='same', activation = 'relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(dropout_rate))
model.add(BatchNormalization())

model.add(Flatten())

model.add(Dense(128, activation = 'relu'))
model.add(Dropout(dropout_rate))

model.add(Dense(62, activation = 'softmax'))

最后记得输出62个神经元进行softmax。

opt = keras.optimizers.rmsprop(lr=0.001, decay=1e-6)
model.compile(optimizer = opt ,
             loss = 'categorical_crossentropy',
             metrics = ['accuracy'])
model.summary()

模型是这样子的：

Model: "sequential_9"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_20 (Conv2D)           (None, 64, 64, 16)        448       
_________________________________________________________________
max_pooling2d_16 (MaxPooling (None, 32, 32, 16)        0         
_________________________________________________________________
dropout_22 (Dropout)         (None, 32, 32, 16)        0         
_________________________________________________________________
batch_normalization_16 (Batc (None, 32, 32, 16)        64        
_________________________________________________________________
conv2d_21 (Conv2D)           (None, 32, 32, 32)        4640      
_________________________________________________________________
max_pooling2d_17 (MaxPooling (None, 16, 16, 32)        0         
_________________________________________________________________
dropout_23 (Dropout)         (None, 16, 16, 32)        0         
_________________________________________________________________
batch_normalization_17 (Batc (None, 16, 16, 32)        128       
_________________________________________________________________
conv2d_22 (Conv2D)           (None, 16, 16, 64)        18496     
_________________________________________________________________
max_pooling2d_18 (MaxPooling (None, 8, 8, 64)          0         
_________________________________________________________________
dropout_24 (Dropout)         (None, 8, 8, 64)          0         
_________________________________________________________________
batch_normalization_18 (Batc (None, 8, 8, 64)          256       
_________________________________________________________________
conv2d_23 (Conv2D)           (None, 8, 8, 128)         73856     
_________________________________________________________________
max_pooling2d_19 (MaxPooling (None, 4, 4, 128)         0         
_________________________________________________________________
dropout_25 (Dropout)         (None, 4, 4, 128)         0         
_________________________________________________________________
batch_normalization_19 (Batc (None, 4, 4, 128)         512       
_________________________________________________________________
conv2d_24 (Conv2D)           (None, 4, 4, 256)         295168    
_________________________________________________________________
max_pooling2d_20 (MaxPooling (None, 2, 2, 256)         0         
_________________________________________________________________
dropout_26 (Dropout)         (None, 2, 2, 256)         0         
_________________________________________________________________
batch_normalization_20 (Batc (None, 2, 2, 256)         1024      
_________________________________________________________________
flatten_7 (Flatten)          (None, 1024)              0         
_________________________________________________________________
dense_13 (Dense)             (None, 128)               131200    
_________________________________________________________________
dropout_27 (Dropout)         (None, 128)               0         
_________________________________________________________________
dense_14 (Dense)             (None, 62)                7998      
=================================================================
Total params: 533,790
Trainable params: 532,798
Non-trainable params: 992

开始训练并保存模型：

history = model.fit(features_train, labels_train, batch_size = batch_size, epochs = epochs,
                    validation_data = [features_test, labels_test])
print("Saving model to disk \n")
####save model for future use
model.save('model.h5')

我们还可以用matplotlib将loss和accuracy刻画出来：

plt.figure(figsize=(10, 5))
plt.subplot(1,2,1)
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title('Model accuracy')
plt.ylabel('Accuracy')
plt.xlabel('Epoch')
plt.legend(['Train', 'Test'], loc='upper left')
####################################
plt.subplot(1,2,2)
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('Model loss')
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.legend(['Train', 'Test'], loc='upper left')
plt.show()

可以看到这个模型的准确率达到了90%以上。

Epoch 20/20
4575/4575 [==============================] - 22s 5ms/step - loss: 0.1753 - acc: 0.9451 - val_loss: 0.2111 - val_acc: 0.9373

随手用手机拍了一张楼外的交通标志并稍微裁剪一下：

使用这张图片之前，我并不知道请勿驶入的标志在哪个文件夹中。

x = Image.open("./DoNotEnter.jpg")
x = x.resize((64, 64), Image.ANTIALIAS)
test = np.array(x)
test = np.expand_dims(test, axis=0)  
print(test.shape)

记住，在测试单张图片之前，必须要扩张成四维张量，而不是直接使用传统的rgb三维矩阵。

predict_test = model.predict_classes(test).astype('int')
print(predict_test)

结果如下：[22]

所以在22号文件夹中放着的是请勿驶入的标志。