Referencia de desempeño, generalización y eficiencia frente al modelo propuesto V1
Se implementó una versión basada en la arquitectura original de EfficientNet-B0, conservando sus bloques MBConv, convoluciones separables en profundidad, mecanismos Squeeze-and-Excitation y conexiones residuales. La capa de clasificación original fue sustituida por una salida sigmoide para realizar clasificación binaria. Este modelo se utilizó como referencia para comparar su desempeño, capacidad de generalización y eficiencia computacional con el modelo propuesto V1.
26,112
Integrado con 1,942 credenciales INE reales (974 anverso, 968 reverso) y enriquecido con 25,138 imágenes externas, como otras identificaciones, para enseñarle al modelo a rechazar documentos que no son una INE.
77.34%
Train (20,197)11.31%
Test (2,955)11.33%
Validate (2,960)300 × 200 px RGB
RMSprop lr=1e-4
Binary Crossentropy
30
| Capa / Orden | Tipo de capa | Filtros / Neuronas | Kernel | Activación | Formato de salida |
|---|---|---|---|---|---|
| Entrada | Input Tensor | Canales RGB | — | — | 200 × 300 × 3 |
| normalizacion | Rescaling (normalización) | — | — | — | 200 × 300 × 3 |
| stem_conv | Conv2D | 32 filtros | (3,3) | — | 100 × 150 × 32 |
| stem_bn | BatchNormalization | — | — | — | 100 × 150 × 32 |
| stem_activation | Activación | — | — | SiLU | 100 × 150 × 32 |
| stage2_block1_depthwise_conv | DepthwiseConv2D | depthwise | (3,3) | — | 100 × 150 × 32 |
| stage2_block1_depthwise_bn | BatchNormalization | — | — | — | 100 × 150 × 32 |
| stage2_block1_depthwise_activation | Activación | — | — | SiLU | 100 × 150 × 32 |
| stage2_block1_se_squeeze | GlobalAveragePooling2D | — | — | — | 32 elementos |
| stage2_block1_se_reshape | Reshape | — | — | — | 1 × 1 × 32 |
| stage2_block1_se_reduce | Conv2D | 8 filtros | (1,1) | — | 1 × 1 × 8 |
| stage2_block1_se_expand | Conv2D | 32 filtros | (1,1) | — | 1 × 1 × 32 |
| stage2_block1_se_multiply | Multiply | — | — | — | 100 × 150 × 32 |
| stage2_block1_project_conv | Conv2D | 16 filtros | (1,1) | — | 100 × 150 × 16 |
| stage2_block1_project_bn | BatchNormalization | — | — | — | 100 × 150 × 16 |
| stage3_block1_expand_conv | Conv2D | 96 filtros | (1,1) | — | 100 × 150 × 96 |
| stage3_block1_expand_bn | BatchNormalization | — | — | — | 100 × 150 × 96 |
| stage3_block1_expand_activation | Activación | — | — | SiLU | 100 × 150 × 96 |
| stage3_block1_depthwise_conv | DepthwiseConv2D | depthwise | (3,3) | — | 50 × 75 × 96 |
| stage3_block1_depthwise_bn | BatchNormalization | — | — | — | 50 × 75 × 96 |
| stage3_block1_depthwise_activation | Activación | — | — | SiLU | 50 × 75 × 96 |
| stage3_block1_se_squeeze | GlobalAveragePooling2D | — | — | — | 96 elementos |
| stage3_block1_se_reshape | Reshape | — | — | — | 1 × 1 × 96 |
| stage3_block1_se_reduce | Conv2D | 4 filtros | (1,1) | — | 1 × 1 × 4 |
| stage3_block1_se_expand | Conv2D | 96 filtros | (1,1) | — | 1 × 1 × 96 |
| stage3_block1_se_multiply | Multiply | — | — | — | 50 × 75 × 96 |
| stage3_block1_project_conv | Conv2D | 24 filtros | (1,1) | — | 50 × 75 × 24 |
| stage3_block1_project_bn | BatchNormalization | — | — | — | 50 × 75 × 24 |
| stage3_block2_expand_conv | Conv2D | 144 filtros | (1,1) | — | 50 × 75 × 144 |
| stage3_block2_expand_bn | BatchNormalization | — | — | — | 50 × 75 × 144 |
| stage3_block2_expand_activation | Activación | — | — | SiLU | 50 × 75 × 144 |
| stage3_block2_depthwise_conv | DepthwiseConv2D | depthwise | (3,3) | — | 50 × 75 × 144 |
| stage3_block2_depthwise_bn | BatchNormalization | — | — | — | 50 × 75 × 144 |
| stage3_block2_depthwise_activation | Activación | — | — | SiLU | 50 × 75 × 144 |
| stage3_block2_se_squeeze | GlobalAveragePooling2D | — | — | — | 144 elementos |
| stage3_block2_se_reshape | Reshape | — | — | — | 1 × 1 × 144 |
| stage3_block2_se_reduce | Conv2D | 6 filtros | (1,1) | — | 1 × 1 × 6 |
| stage3_block2_se_expand | Conv2D | 144 filtros | (1,1) | — | 1 × 1 × 144 |
| stage3_block2_se_multiply | Multiply | — | — | — | 50 × 75 × 144 |
| stage3_block2_project_conv | Conv2D | 24 filtros | (1,1) | — | 50 × 75 × 24 |
| stage3_block2_project_bn | BatchNormalization | — | — | — | 50 × 75 × 24 |
| stage3_block2_drop_connect | Dropout (Regularización) | rate = 0.025 | — | — | 24 elementos |
| stage3_block2_residual | Add (conexión residual) | — | — | — | 50 × 75 × 24 |
| stage4_block1_expand_conv | Conv2D | 144 filtros | (1,1) | — | 50 × 75 × 144 |
| stage4_block1_expand_bn | BatchNormalization | — | — | — | 50 × 75 × 144 |
| stage4_block1_expand_activation | Activación | — | — | SiLU | 50 × 75 × 144 |
| stage4_block1_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 25 × 38 × 144 |
| stage4_block1_depthwise_bn | BatchNormalization | — | — | — | 25 × 38 × 144 |
| stage4_block1_depthwise_activation | Activación | — | — | SiLU | 25 × 38 × 144 |
| stage4_block1_se_squeeze | GlobalAveragePooling2D | — | — | — | 144 elementos |
| stage4_block1_se_reshape | Reshape | — | — | — | 1 × 1 × 144 |
| stage4_block1_se_reduce | Conv2D | 6 filtros | (1,1) | — | 1 × 1 × 6 |
| stage4_block1_se_expand | Conv2D | 144 filtros | (1,1) | — | 1 × 1 × 144 |
| stage4_block1_se_multiply | Multiply | — | — | — | 25 × 38 × 144 |
| stage4_block1_project_conv | Conv2D | 40 filtros | (1,1) | — | 25 × 38 × 40 |
| stage4_block1_project_bn | BatchNormalization | — | — | — | 25 × 38 × 40 |
| stage4_block2_expand_conv | Conv2D | 240 filtros | (1,1) | — | 25 × 38 × 240 |
| stage4_block2_expand_bn | BatchNormalization | — | — | — | 25 × 38 × 240 |
| stage4_block2_expand_activation | Activación | — | — | SiLU | 25 × 38 × 240 |
| stage4_block2_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 25 × 38 × 240 |
| stage4_block2_depthwise_bn | BatchNormalization | — | — | — | 25 × 38 × 240 |
| stage4_block2_depthwise_activation | Activación | — | — | SiLU | 25 × 38 × 240 |
| stage4_block2_se_squeeze | GlobalAveragePooling2D | — | — | — | 240 elementos |
| stage4_block2_se_reshape | Reshape | — | — | — | 1 × 1 × 240 |
| stage4_block2_se_reduce | Conv2D | 10 filtros | (1,1) | — | 1 × 1 × 10 |
| stage4_block2_se_expand | Conv2D | 240 filtros | (1,1) | — | 1 × 1 × 240 |
| stage4_block2_se_multiply | Multiply | — | — | — | 25 × 38 × 240 |
| stage4_block2_project_conv | Conv2D | 40 filtros | (1,1) | — | 25 × 38 × 40 |
| stage4_block2_project_bn | BatchNormalization | — | — | — | 25 × 38 × 40 |
| stage4_block2_drop_connect | Dropout (Regularización) | rate = 0.05 | — | — | 40 elementos |
| stage4_block2_residual | Add (conexión residual) | — | — | — | 25 × 38 × 40 |
| stage5_block1_expand_conv | Conv2D | 240 filtros | (1,1) | — | 25 × 38 × 240 |
| stage5_block1_expand_bn | BatchNormalization | — | — | — | 25 × 38 × 240 |
| stage5_block1_expand_activation | Activación | — | — | SiLU | 25 × 38 × 240 |
| stage5_block1_depthwise_conv | DepthwiseConv2D | depthwise | (3,3) | — | 13 × 19 × 240 |
| stage5_block1_depthwise_bn | BatchNormalization | — | — | — | 13 × 19 × 240 |
| stage5_block1_depthwise_activation | Activación | — | — | SiLU | 13 × 19 × 240 |
| stage5_block1_se_squeeze | GlobalAveragePooling2D | — | — | — | 240 elementos |
| stage5_block1_se_reshape | Reshape | — | — | — | 1 × 1 × 240 |
| stage5_block1_se_reduce | Conv2D | 10 filtros | (1,1) | — | 1 × 1 × 10 |
| stage5_block1_se_expand | Conv2D | 240 filtros | (1,1) | — | 1 × 1 × 240 |
| stage5_block1_se_multiply | Multiply | — | — | — | 13 × 19 × 240 |
| stage5_block1_project_conv | Conv2D | 80 filtros | (1,1) | — | 13 × 19 × 80 |
| stage5_block1_project_bn | BatchNormalization | — | — | — | 13 × 19 × 80 |
| stage5_block2_expand_conv | Conv2D | 480 filtros | (1,1) | — | 13 × 19 × 480 |
| stage5_block2_expand_bn | BatchNormalization | — | — | — | 13 × 19 × 480 |
| stage5_block2_expand_activation | Activación | — | — | SiLU | 13 × 19 × 480 |
| stage5_block2_depthwise_conv | DepthwiseConv2D | depthwise | (3,3) | — | 13 × 19 × 480 |
| stage5_block2_depthwise_bn | BatchNormalization | — | — | — | 13 × 19 × 480 |
| stage5_block2_depthwise_activation | Activación | — | — | SiLU | 13 × 19 × 480 |
| stage5_block2_se_squeeze | GlobalAveragePooling2D | — | — | — | 480 elementos |
| stage5_block2_se_reshape | Reshape | — | — | — | 1 × 1 × 480 |
| stage5_block2_se_reduce | Conv2D | 20 filtros | (1,1) | — | 1 × 1 × 20 |
| stage5_block2_se_expand | Conv2D | 480 filtros | (1,1) | — | 1 × 1 × 480 |
| stage5_block2_se_multiply | Multiply | — | — | — | 13 × 19 × 480 |
| stage5_block2_project_conv | Conv2D | 80 filtros | (1,1) | — | 13 × 19 × 80 |
| stage5_block2_project_bn | BatchNormalization | — | — | — | 13 × 19 × 80 |
| stage5_block2_drop_connect | Dropout (Regularización) | rate = 0.07500000000000001 | — | — | 80 elementos |
| stage5_block2_residual | Add (conexión residual) | — | — | — | 13 × 19 × 80 |
| stage5_block3_expand_conv | Conv2D | 480 filtros | (1,1) | — | 13 × 19 × 480 |
| stage5_block3_expand_bn | BatchNormalization | — | — | — | 13 × 19 × 480 |
| stage5_block3_expand_activation | Activación | — | — | SiLU | 13 × 19 × 480 |
| stage5_block3_depthwise_conv | DepthwiseConv2D | depthwise | (3,3) | — | 13 × 19 × 480 |
| stage5_block3_depthwise_bn | BatchNormalization | — | — | — | 13 × 19 × 480 |
| stage5_block3_depthwise_activation | Activación | — | — | SiLU | 13 × 19 × 480 |
| stage5_block3_se_squeeze | GlobalAveragePooling2D | — | — | — | 480 elementos |
| stage5_block3_se_reshape | Reshape | — | — | — | 1 × 1 × 480 |
| stage5_block3_se_reduce | Conv2D | 20 filtros | (1,1) | — | 1 × 1 × 20 |
| stage5_block3_se_expand | Conv2D | 480 filtros | (1,1) | — | 1 × 1 × 480 |
| stage5_block3_se_multiply | Multiply | — | — | — | 13 × 19 × 480 |
| stage5_block3_project_conv | Conv2D | 80 filtros | (1,1) | — | 13 × 19 × 80 |
| stage5_block3_project_bn | BatchNormalization | — | — | — | 13 × 19 × 80 |
| stage5_block3_drop_connect | Dropout (Regularización) | rate = 0.08750000000000001 | — | — | 80 elementos |
| stage5_block3_residual | Add (conexión residual) | — | — | — | 13 × 19 × 80 |
| stage6_block1_expand_conv | Conv2D | 480 filtros | (1,1) | — | 13 × 19 × 480 |
| stage6_block1_expand_bn | BatchNormalization | — | — | — | 13 × 19 × 480 |
| stage6_block1_expand_activation | Activación | — | — | SiLU | 13 × 19 × 480 |
| stage6_block1_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 13 × 19 × 480 |
| stage6_block1_depthwise_bn | BatchNormalization | — | — | — | 13 × 19 × 480 |
| stage6_block1_depthwise_activation | Activación | — | — | SiLU | 13 × 19 × 480 |
| stage6_block1_se_squeeze | GlobalAveragePooling2D | — | — | — | 480 elementos |
| stage6_block1_se_reshape | Reshape | — | — | — | 1 × 1 × 480 |
| stage6_block1_se_reduce | Conv2D | 20 filtros | (1,1) | — | 1 × 1 × 20 |
| stage6_block1_se_expand | Conv2D | 480 filtros | (1,1) | — | 1 × 1 × 480 |
| stage6_block1_se_multiply | Multiply | — | — | — | 13 × 19 × 480 |
| stage6_block1_project_conv | Conv2D | 112 filtros | (1,1) | — | 13 × 19 × 112 |
| stage6_block1_project_bn | BatchNormalization | — | — | — | 13 × 19 × 112 |
| stage6_block2_expand_conv | Conv2D | 672 filtros | (1,1) | — | 13 × 19 × 672 |
| stage6_block2_expand_bn | BatchNormalization | — | — | — | 13 × 19 × 672 |
| stage6_block2_expand_activation | Activación | — | — | SiLU | 13 × 19 × 672 |
| stage6_block2_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 13 × 19 × 672 |
| stage6_block2_depthwise_bn | BatchNormalization | — | — | — | 13 × 19 × 672 |
| stage6_block2_depthwise_activation | Activación | — | — | SiLU | 13 × 19 × 672 |
| stage6_block2_se_squeeze | GlobalAveragePooling2D | — | — | — | 672 elementos |
| stage6_block2_se_reshape | Reshape | — | — | — | 1 × 1 × 672 |
| stage6_block2_se_reduce | Conv2D | 28 filtros | (1,1) | — | 1 × 1 × 28 |
| stage6_block2_se_expand | Conv2D | 672 filtros | (1,1) | — | 1 × 1 × 672 |
| stage6_block2_se_multiply | Multiply | — | — | — | 13 × 19 × 672 |
| stage6_block2_project_conv | Conv2D | 112 filtros | (1,1) | — | 13 × 19 × 112 |
| stage6_block2_project_bn | BatchNormalization | — | — | — | 13 × 19 × 112 |
| stage6_block2_drop_connect | Dropout (Regularización) | rate = 0.1125 | — | — | 112 elementos |
| stage6_block2_residual | Add (conexión residual) | — | — | — | 13 × 19 × 112 |
| stage6_block3_expand_conv | Conv2D | 672 filtros | (1,1) | — | 13 × 19 × 672 |
| stage6_block3_expand_bn | BatchNormalization | — | — | — | 13 × 19 × 672 |
| stage6_block3_expand_activation | Activación | — | — | SiLU | 13 × 19 × 672 |
| stage6_block3_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 13 × 19 × 672 |
| stage6_block3_depthwise_bn | BatchNormalization | — | — | — | 13 × 19 × 672 |
| stage6_block3_depthwise_activation | Activación | — | — | SiLU | 13 × 19 × 672 |
| stage6_block3_se_squeeze | GlobalAveragePooling2D | — | — | — | 672 elementos |
| stage6_block3_se_reshape | Reshape | — | — | — | 1 × 1 × 672 |
| stage6_block3_se_reduce | Conv2D | 28 filtros | (1,1) | — | 1 × 1 × 28 |
| stage6_block3_se_expand | Conv2D | 672 filtros | (1,1) | — | 1 × 1 × 672 |
| stage6_block3_se_multiply | Multiply | — | — | — | 13 × 19 × 672 |
| stage6_block3_project_conv | Conv2D | 112 filtros | (1,1) | — | 13 × 19 × 112 |
| stage6_block3_project_bn | BatchNormalization | — | — | — | 13 × 19 × 112 |
| stage6_block3_drop_connect | Dropout (Regularización) | rate = 0.125 | — | — | 112 elementos |
| stage6_block3_residual | Add (conexión residual) | — | — | — | 13 × 19 × 112 |
| stage7_block1_expand_conv | Conv2D | 672 filtros | (1,1) | — | 13 × 19 × 672 |
| stage7_block1_expand_bn | BatchNormalization | — | — | — | 13 × 19 × 672 |
| stage7_block1_expand_activation | Activación | — | — | SiLU | 13 × 19 × 672 |
| stage7_block1_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 7 × 10 × 672 |
| stage7_block1_depthwise_bn | BatchNormalization | — | — | — | 7 × 10 × 672 |
| stage7_block1_depthwise_activation | Activación | — | — | SiLU | 7 × 10 × 672 |
| stage7_block1_se_squeeze | GlobalAveragePooling2D | — | — | — | 672 elementos |
| stage7_block1_se_reshape | Reshape | — | — | — | 1 × 1 × 672 |
| stage7_block1_se_reduce | Conv2D | 28 filtros | (1,1) | — | 1 × 1 × 28 |
| stage7_block1_se_expand | Conv2D | 672 filtros | (1,1) | — | 1 × 1 × 672 |
| stage7_block1_se_multiply | Multiply | — | — | — | 7 × 10 × 672 |
| stage7_block1_project_conv | Conv2D | 192 filtros | (1,1) | — | 7 × 10 × 192 |
| stage7_block1_project_bn | BatchNormalization | — | — | — | 7 × 10 × 192 |
| stage7_block2_expand_conv | Conv2D | 1152 filtros | (1,1) | — | 7 × 10 × 1152 |
| stage7_block2_expand_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage7_block2_expand_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage7_block2_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 7 × 10 × 1152 |
| stage7_block2_depthwise_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage7_block2_depthwise_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage7_block2_se_squeeze | GlobalAveragePooling2D | — | — | — | 1,152 elementos |
| stage7_block2_se_reshape | Reshape | — | — | — | 1 × 1 × 1152 |
| stage7_block2_se_reduce | Conv2D | 48 filtros | (1,1) | — | 1 × 1 × 48 |
| stage7_block2_se_expand | Conv2D | 1152 filtros | (1,1) | — | 1 × 1 × 1152 |
| stage7_block2_se_multiply | Multiply | — | — | — | 7 × 10 × 1152 |
| stage7_block2_project_conv | Conv2D | 192 filtros | (1,1) | — | 7 × 10 × 192 |
| stage7_block2_project_bn | BatchNormalization | — | — | — | 7 × 10 × 192 |
| stage7_block2_drop_connect | Dropout (Regularización) | rate = 0.15000000000000002 | — | — | 192 elementos |
| stage7_block2_residual | Add (conexión residual) | — | — | — | 7 × 10 × 192 |
| stage7_block3_expand_conv | Conv2D | 1152 filtros | (1,1) | — | 7 × 10 × 1152 |
| stage7_block3_expand_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage7_block3_expand_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage7_block3_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 7 × 10 × 1152 |
| stage7_block3_depthwise_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage7_block3_depthwise_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage7_block3_se_squeeze | GlobalAveragePooling2D | — | — | — | 1,152 elementos |
| stage7_block3_se_reshape | Reshape | — | — | — | 1 × 1 × 1152 |
| stage7_block3_se_reduce | Conv2D | 48 filtros | (1,1) | — | 1 × 1 × 48 |
| stage7_block3_se_expand | Conv2D | 1152 filtros | (1,1) | — | 1 × 1 × 1152 |
| stage7_block3_se_multiply | Multiply | — | — | — | 7 × 10 × 1152 |
| stage7_block3_project_conv | Conv2D | 192 filtros | (1,1) | — | 7 × 10 × 192 |
| stage7_block3_project_bn | BatchNormalization | — | — | — | 7 × 10 × 192 |
| stage7_block3_drop_connect | Dropout (Regularización) | rate = 0.1625 | — | — | 192 elementos |
| stage7_block3_residual | Add (conexión residual) | — | — | — | 7 × 10 × 192 |
| stage7_block4_expand_conv | Conv2D | 1152 filtros | (1,1) | — | 7 × 10 × 1152 |
| stage7_block4_expand_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage7_block4_expand_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage7_block4_depthwise_conv | DepthwiseConv2D | depthwise | (5,5) | — | 7 × 10 × 1152 |
| stage7_block4_depthwise_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage7_block4_depthwise_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage7_block4_se_squeeze | GlobalAveragePooling2D | — | — | — | 1,152 elementos |
| stage7_block4_se_reshape | Reshape | — | — | — | 1 × 1 × 1152 |
| stage7_block4_se_reduce | Conv2D | 48 filtros | (1,1) | — | 1 × 1 × 48 |
| stage7_block4_se_expand | Conv2D | 1152 filtros | (1,1) | — | 1 × 1 × 1152 |
| stage7_block4_se_multiply | Multiply | — | — | — | 7 × 10 × 1152 |
| stage7_block4_project_conv | Conv2D | 192 filtros | (1,1) | — | 7 × 10 × 192 |
| stage7_block4_project_bn | BatchNormalization | — | — | — | 7 × 10 × 192 |
| stage7_block4_drop_connect | Dropout (Regularización) | rate = 0.17500000000000002 | — | — | 192 elementos |
| stage7_block4_residual | Add (conexión residual) | — | — | — | 7 × 10 × 192 |
| stage8_block1_expand_conv | Conv2D | 1152 filtros | (1,1) | — | 7 × 10 × 1152 |
| stage8_block1_expand_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage8_block1_expand_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage8_block1_depthwise_conv | DepthwiseConv2D | depthwise | (3,3) | — | 7 × 10 × 1152 |
| stage8_block1_depthwise_bn | BatchNormalization | — | — | — | 7 × 10 × 1152 |
| stage8_block1_depthwise_activation | Activación | — | — | SiLU | 7 × 10 × 1152 |
| stage8_block1_se_squeeze | GlobalAveragePooling2D | — | — | — | 1,152 elementos |
| stage8_block1_se_reshape | Reshape | — | — | — | 1 × 1 × 1152 |
| stage8_block1_se_reduce | Conv2D | 48 filtros | (1,1) | — | 1 × 1 × 48 |
| stage8_block1_se_expand | Conv2D | 1152 filtros | (1,1) | — | 1 × 1 × 1152 |
| stage8_block1_se_multiply | Multiply | — | — | — | 7 × 10 × 1152 |
| stage8_block1_project_conv | Conv2D | 320 filtros | (1,1) | — | 7 × 10 × 320 |
| stage8_block1_project_bn | BatchNormalization | — | — | — | 7 × 10 × 320 |
| head_conv | Conv2D | 1280 filtros | (1,1) | — | 7 × 10 × 1280 |
| head_bn | BatchNormalization | — | — | — | 7 × 10 × 1280 |
| head_activation | Activación | — | — | SiLU | 7 × 10 × 1280 |
| global_average_pooling | GlobalAveragePooling2D | — | — | — | 1,280 elementos |
| classifier_dropout | Dropout (Regularización) | rate = 0.2 | — | — | 1,280 elementos |
| Salida | Dense (Clasificador binario) | 1 neuronas | — | Sigmoide | 1 (Probabilidad de INE) |
Verificación de Identificaciones mediante Redes Neuronales
Pérdida
Precisión
144
7
1
2,822
V1 (MFHA_CDO, el modelo propuesto) contra tres arquitecturas de referencia entrenadas sobre el mismo dataset — MobileNetV2, SqueezeNet y EfficientNet-B0.
V1 · MFHA_CDO
V0.1 · MobileNetV2
V0.2 · SqueezeNet
V0.3 · EfficientNet-B0
Tras 30 épocas: 99.6% train · 99.9% val
Tras 30 épocas: 99.65% train · 99.2% val
Tras 30 épocas: 97.45% train · 95.9% val
Tras 30 épocas: 99.55% train · 99.8% val
Global 99.66%
2964/2974 aciertosBalanceada 99.20%INE 98.68% · 149/151Global 99.43%
2957/2974 aciertosBalanceada 95.31%INE 90.73% · 137/151Global 94.92%
2823/2974 aciertosBalanceada 50.00%INE 0.00% · 0/151Global 99.73%
2966/2974 aciertosBalanceada 97.66%INE 95.36% · 144/15120.47 MB
4.44 MB (4.6× más chico que V1)
1.43 MB (14.3× más chico que V1)
7.90 MB (2.6× más chico que V1)
Referencia (V1)
~4.6× más rápido que V1
~14.3× más rápido que V1
~2.6× más rápido que V1