第6期 汤礼颖，等：一种卷积神经网络集成的多样性度量方法 ·1033· 续表2 结构 Model 2 Model 3 Model 5 Model 6 卷积层2 Conv3-1283×3 Conv3-2563×3 Conw3-1283×3 Conv3-2563×3 卷积层3 池化层1 Maxpool Maxpool Maxpool Maxpool 卷积层4 Conv3-2563×3 Conv3-128 3x3 Conv3-2563×3 Conv3-1283×3 卷积层5 Conv3-5123×3 Conv3-643×3 Conv3-5123×3 Conv3-643×3 卷积层6 Conv3-643×3 Conv3-5123×3 池化层2 Maxpool Maxpool Maxpool Maxpool 过拟合层1 Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) 卷积层7 Conv3-1283×3 Conw3-5123×3 Conv3-2563×3 Conv3-643×3 卷积层8 Conv3-2563×3 Conw3-2563×3 Conv3-1283×3 Conv3-1283×3 卷积层9 Conv3-5123×3 Conv3-643×3 Conv3-2563×3 池化层3 Maxpool 卷积层10 Conv3-1283×3 Conv3-5123×3 卷积层11 Conv3-64 3x3 卷积层12 池化层4 Global Maxpool Global Maxpool Global Maxpool Global Maxpool 拉平 fatten fatten fatten fatten 过拟合层2 Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) 全连接层 Dense(64) Dense(65) Dense(66) Dense(67) 输出层 Softmax Softmax Softmax Softmax 表35个候选基模型结构 Table 3 Structures of five-candidate basic models 结构 Model 8 Model 10 Model 12 Model_13 Model 14 输入层 32×32×3 32×32×3 32×32×3 32×32×3 32×32×3 卷积层1 Conv3-2563×3 Conv3-5123×3 Conv3-643×3 Conv3-641×1 Conv3-643×3 卷积层2 Conw3-1283×3 Conv3-64 3x3 Conw3-641×1 Conv.3-643×3 Conv3-641×3 卷积层3 Conv3-641×1 Conv3-643×1 池化层1 Maxpool Maxpool Maxpool Maxpool Maxpool 卷积层4 Conv3-2563×3 Conv3-1283×3 Conv3-2563×3 Conw3-2561×1 Conv3-128 3x3 卷积层5 Conv3-1283×3 Conv3-643×3 Conv3-2561×1 Conv3-2563×3 Conv3-1281×3 卷积层6 Conw3-2563×3 Conw3-2561×1 Conv3-l283×1 池化层2 Maxpool Maxpool Maxpool Maxpool Maxpool 过拟合层1 Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) 卷积层7 Conv3-1283×3 Conv3-5123×3 Conv3-5123×3 Conw3-5121×1 Conv3-2563×3 卷积层8 Conv3-2563×3 Conv3-643×3 Conv3-5121×1 Conv3-5123×3 Conv3-2561×3 卷积层9 Conv3-643×3 Conv3-5121×1 Conv3-2563×1 池化层3 Maxpool Maxpool Maxpool Maxpool 卷积层10 Conv3-1283×3 Conv3-1283×3 Conw3-1281×1 Conv3-5123×3 卷积层11 Conv3-64 3x3 Conv3-1281×1 Conw3-1283×3 Conv3-5121×3 卷积层12 Conw3-1281×1 Conv3-5123×1 池化层4 Global Maxpool Global Maxpool Global Maxpool Global Maxpool GlobalMaxpool续表 2 结构 Model_2 Model_3 Model_5 Model_6 卷积层2 Conv3-128 3×3 Conv3-256 3×3 Conv3-128 3×3 Conv3-256 3×3 卷积层3 — — — — 池化层1 Maxpool Maxpool Maxpool Maxpool 卷积层4 Conv3-256 3×3 Conv3-128 3×3 Conv3-256 3×3 Conv3-128 3×3 卷积层5 Conv3-512 3×3 Conv3-64 3×3 Conv3-512 3×3 Conv3-64 3×3 卷积层6 Conv3-64 3×3 — Conv3-512 3×3 — 池化层2 Maxpool Maxpool Maxpool Maxpool 过拟合层1 Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) 卷积层7 Conv3-128 3×3 Conv3-512 3×3 Conv3-256 3×3 Conv3-64 3×3 卷积层8 Conv3-256 3×3 Conv3-256 3×3 Conv3-128 3×3 Conv3-128 3×3 卷积层9 Conv3-512 3×3 — Conv3-64 3×3 Conv3-256 3×3 池化层3 — Maxpool — — 卷积层10 — Conv3-128 3×3 — Conv3-512 3×3 卷积层11 — Conv3-64 3×3 — 卷积层12 — — — — 池化层4 Global Maxpool Global Maxpool Global Maxpool Global Maxpool 拉平 fatten fatten fatten fatten 过拟合层2 Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) 全连接层 Dense(64) Dense(65) Dense(66) Dense(67) 输出层 Softmax Softmax Softmax Softmax 表 3 5 个候选基模型结构 Table 3 Structures of five-candidate basic models 结构 Model_8 Model_10 Model_12 Model_13 Model_14 输入层 32×32 ×3 32×32 ×3 32×32 ×3 32×32 ×3 32×32 ×3 卷积层1 Conv3-256 3×3 Conv3-512 3×3 Conv3-64 3×3 Conv3-64 1×1 Conv3-64 3×3 卷积层2 Conv3-128 3×3 Conv3-64 3×3 Conv3-64 1×1 Conv3-64 3×3 Conv3-64 1×3 卷积层3 — — — Conv3-64 1×1 Conv3-64 3×1 池化层1 Maxpool Maxpool Maxpool Maxpool Maxpool 卷积层4 Conv3-256 3×3 Conv3-128 3×3 Conv3-256 3×3 Conv3-256 1×1 Conv3-128 3×3 卷积层5 Conv3-128 3×3 Conv3-64 3×3 Conv3-256 1×1 Conv3-2563×3 Conv3-128 1×3 卷积层6 Conv3-256 3×3 — — Conv3-256 1×1 Conv3-128 3×1 池化层2 Maxpool Maxpool Maxpool Maxpool Maxpool 过拟合层1 Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) Dropout(0.25) 卷积层7 Conv3-128 3×3 Conv3-512 3×3 Conv3-512 3×3 Conv3-512 1×1 Conv3-256 3×3 卷积层8 Conv3-256 3×3 Conv3-64 3×3 Conv3-512 1×1 Conv3-512 3×3 Conv3-256 1×3 卷积层9 Conv3-64 3×3 — — Conv3-512 1×1 Conv3-256 3×1 池化层3 — Maxpool Maxpool Maxpool Maxpool 卷积层10 — Conv3-128 3×3 Conv3-128 3×3 Conv3-128 1×1 Conv3-512 3×3 卷积层11 — Conv3-64 3×3 Conv3-128 1×1 Conv3-128 3×3 Conv3-512 1×3 卷积层12 — — — Conv3-128 1×1 Conv3-512 3×1 池化层4 Global Maxpool Global Maxpool Global Maxpool Global Maxpool GlobalMaxpool 第 6 期 汤礼颖，等：一种卷积神经网络集成的多样性度量方法 ·1033·