AlexNet,VGG,Inceptions, ResNet, MobileNet对比

以下是几种经典卷积神经网络（CNN）架构的对比与关键特性分析：

AlexNet（2012）

model = Sequential([Conv2D(96, (11,11), strides=4, activation='relu', input_shape=(227,227,3)),MaxPooling2D((3,3), strides=2),Conv2D(256, (5,5), padding='same', activation='relu'),MaxPooling2D((3,3), strides=2),Conv2D(384, (3,3), padding='same', activation='relu'),Conv2D(384, (3,3), padding='same', activation='relu'),Conv2D(256, (3,3), padding='same', activation='relu'),MaxPooling2D((3,3), strides=2),Flatten(),Dense(4096, activation='relu'),Dense(4096, activation='relu'),Dense(1000, activation='softmax')
])

• 首个深层CNN，5个卷积层+3个全连接层
• 使用ReLU激活函数提升训练速度
• 引入Dropout防止过拟合

VGG（2014）

def VGG_block(num_convs, filters):block = Sequential()for _ in range(num_convs):block.add(Conv2D(filters, (3,3), padding='same', activation='relu'))block.add(MaxPooling2D((2,2), strides=2))return block

• 统一使用3x3小卷积核堆叠
• 典型配置VGG16/VGG19（16/19层）
• 参数量大（138M），全连接层占比高

Inception系列（2014-2016）

def Inception_module(x, filters):branch1 = Conv2D(filters[0], (1,1), padding='same', activation='relu')(x)branch2 = Conv2D(filters[1], (1,1), padding='same', activation='relu')(x)branch2 = Conv2D(filters[2], (3,3), padding='same', activation='relu')(branch2)branch3 = Conv2D(filters[3], (1,1), padding='same', activation='relu')(x)branch3 = Conv2D(filters[4], (5,5), padding='same', activation='relu')(branch3)branch4 = MaxPooling2D((3,3), strides=1, padding='same')(x)branch4 = Conv2D(filters[5], (1,1), padding='same', activation='relu')(branch4)return concatenate([branch1, branch2, branch3, branch4])

• 多尺度并行卷积（1x1,3x3,5x5）
• 引入1x1卷积降维减少计算量
• v3加入BN层，v4结合ResNet思想

ResNet（2015）

def residual_block(x, filters, strides=1):shortcut = xx = Conv2D(filters, (3,3), strides=strides, padding='same')(x)x = BatchNormalization()(x)x = ReLU()(x)x = Conv2D(filters, (3,3), padding='same')(x)x = BatchNormalization()(x)if strides != 1 or shortcut.shape[-1] != filters:shortcut = Conv2D(filters, (1,1), strides=strides)(shortcut)shortcut = BatchNormalization()(shortcut)x = Add()([x, shortcut])return ReLU()(x)

• 残差连接解决梯度消失问题
• 典型结构ResNet50/101/152
• 使用Bottleneck结构（1x1-3x3-1x1）

MobileNet（2017）

def depthwise_separable_conv(x, filters, strides):x = DepthwiseConv2D((3,3), strides=strides, padding='same')(x)x = BatchNormalization()(x)x = ReLU()(x)x = Conv2D(filters, (1,1), padding='same')(x)x = BatchNormalization()(x)return ReLU()(x)

• 深度可分离卷积降低计算量
• 宽度乘子调整模型大小
• v2引入倒残差结构和线性瓶颈
• v3加入SE模块和h-swish激活

这些架构的发展趋势表现为：网络深度增加（AlexNet 8层 → ResNet 152层）、计算效率提升（标准卷积 → 深度可分离卷积）、模块化设计（Inception/Residual块）、准确率与速度的平衡（MobileNet系列）。实际应用中需根据计算资源和精度要求选择合适的架构。