如何改进VGG16分类模型，测试数据集的准确率仅为60％

Question

我正在尝试对屋顶瓦片进行 16 类分类。我的原始训练数据集总共包含 6800 张图像。我使用离线增强来为每个类生成额外的 750 张图像。 测试数据集是在拍摄屋顶瓦片图像的单独会话中收集的，总共包含 1100 张图像。我在这里也使用了离线增强，并从该数据集中为每个类生成了额外的 450 张图像。
该模型在训练数据集上的准确率为 93%，在验证集上的准确率为 84.2%。然而，我在测试数据集上获得的最大准确度是 60%。
我尝试在模型中添加 dropout 层、更改批量大小等，但测试集的准确率没有超过 60%。我错过了什么？

import pathlib

data_dir = pathlib.Path("C:/Users/../train_750Aug")

train_data_dir      = data_dir
validation_data_dir = data_dir

test_data_dir = pathlib.Path("C:/Users/../Testset_cubed_aug450")
nr_of_epochs  = 20
saved_model_name = 'C:/Users/../models/750augTrain_Aug450Test.h5'


batch_size = 32
img_height = 224
img_width =  224

datagen=ImageDataGenerator(rescale= 1. /255,validation_split=0.25,width_shift_range=[-10,10],horizontal_flip=True
                               ,brightness_range=[0.5,1.5],vertical_flip=True)


train_generator=datagen.flow_from_directory(data_dir,
                                                target_size=(img_width,img_height), 
                                                batch_size=batch_size, 
                                                subset="training",
                                                class_mode='categorical'
                                               )
validation_generator=datagen.flow_from_directory(data_dir, 
                                            target_size=(img_width,img_height), 
                                            batch_size=batch_size, 
                                            subset="validation",
                                            class_mode='categorical'
                                           )

test_datagen = ImageDataGenerator(rescale= 1. /255)
test_generator = test_datagen.flow_from_directory(test_data_dir,
                                           target_size = (img_width,img_height),
                                           batch_size = batch_size,
                                           class_mode = 'categorical')




def define_VGGmodel():
    # load model
    model = VGG16(include_top=False, input_shape=(img_width, img_height,3))
    # mark loaded layers as not trainable
    list_trainable = ['block5_conv3'] 
    for layer in model.layers:
        if(layer.name in list_trainable):
            layer.trainable = True
            print(layer.name)
        else:
            layer.trainable = False
        
    ## make layers trainable 
    
    # add new classifier layers
    flat1 = Flatten()(model.layers[-1].output)
    class1 = Dense(128, activation='relu', kernel_initializer='he_uniform')(flat1)
    droppyDrop = Dropout(0.2)(class1)
       
    output = Dense(16, activation='softmax')(droppyDrop)
    # define new model
    model = Model(inputs=model.inputs, outputs=output)
    # compile model    
    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['categorical_accuracy'])
    return model
model1 = define_VGGmodel()

from tensorflow.keras.callbacks import History
history=History()
    history = model1.fit(train_generator,steps_per_epoch=len(train_generator),epochs=nr_of_epochs,validation_data=validation_generator,shuffle=True, verbose=2)

model1.save(saved_model_name)

我尝试对训练和测试数据集进行离线增强，包括更改批量大小、添加 dropout 层，但测试集的准确性没有提高并保持在 60%

Answer 1

您的分类器可能过于受限。你有

1. Dense(128) -> ReLu

2. DropOut(0.2)

3. Dense(16) -> SoftMax

而 VGG16 与

include_top=True

有

1. Dense(4096) -> ReLu

2. Dense(4096) -> ReLu

3. Dense(16) -> SoftMax

我建议在你的分类器中投入更多的权重和层数，看看它会带你做什么。