Verificación de Identificaciones
mediante Redes Neuronales

Desliza para comenzar↓

Comparativa

Referencia ligera y eficiente frente al modelo propuesto V1

MobileNetV2

Se implementó una versión basada en la arquitectura original de MobileNetV2, preservando sus bloques residuales invertidos, cuellos de botella lineales y convoluciones separables en profundidad. La salida multiclase original fue reemplazada por una neurona con activación sigmoide para resolver la tarea de clasificación binaria. Este modelo permitió establecer una referencia ligera y eficiente para comparar los resultados obtenidos por el modelo V1.

Bloques residuales invertidos

Cuellos de botella lineales

Convoluciones separables en profundidad

Salida sigmoide binaria

Dataset

26,112

Imágenes totales

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)

Arquitectura

Entrada

300 × 200 px RGB

Optimizer

RMSprop lr=1e-4

Loss

Binary Crossentropy

Épocas

Desliza la tabla para ver más →

Capa / Orden	Tipo de capa	Filtros / Neuronas	Kernel	Activación	Formato de salida
Entrada	Input Tensor	Canales RGB	—	—	200 × 300 × 3
conv2d	Conv2D	32 filtros	(3,3)	—	100 × 150 × 32
batch_normalization	BatchNormalization	—	—	—	100 × 150 × 32
re_lu	ReLU	—	—	—	100 × 150 × 32
depthwise_conv2d	DepthwiseConv2D	depthwise	(3,3)	—	100 × 150 × 32
batch_normalization_1	BatchNormalization	—	—	—	100 × 150 × 32
re_lu_1	ReLU	—	—	—	100 × 150 × 32
conv2d_1	Conv2D	16 filtros	(1,1)	—	100 × 150 × 16
batch_normalization_2	BatchNormalization	—	—	—	100 × 150 × 16
conv2d_2	Conv2D	96 filtros	(1,1)	—	100 × 150 × 96
batch_normalization_3	BatchNormalization	—	—	—	100 × 150 × 96
re_lu_2	ReLU	—	—	—	100 × 150 × 96
depthwise_conv2d_1	DepthwiseConv2D	depthwise	(3,3)	—	50 × 75 × 96
batch_normalization_4	BatchNormalization	—	—	—	50 × 75 × 96
re_lu_3	ReLU	—	—	—	50 × 75 × 96
conv2d_3	Conv2D	24 filtros	(1,1)	—	50 × 75 × 24
batch_normalization_5	BatchNormalization	—	—	—	50 × 75 × 24
conv2d_4	Conv2D	144 filtros	(1,1)	—	50 × 75 × 144
batch_normalization_6	BatchNormalization	—	—	—	50 × 75 × 144
re_lu_4	ReLU	—	—	—	50 × 75 × 144
depthwise_conv2d_2	DepthwiseConv2D	depthwise	(3,3)	—	50 × 75 × 144
batch_normalization_7	BatchNormalization	—	—	—	50 × 75 × 144
re_lu_5	ReLU	—	—	—	50 × 75 × 144
conv2d_5	Conv2D	24 filtros	(1,1)	—	50 × 75 × 24
batch_normalization_8	BatchNormalization	—	—	—	50 × 75 × 24
add	Add (conexión residual)	—	—	—	50 × 75 × 24
conv2d_6	Conv2D	144 filtros	(1,1)	—	50 × 75 × 144
batch_normalization_9	BatchNormalization	—	—	—	50 × 75 × 144
re_lu_6	ReLU	—	—	—	50 × 75 × 144
depthwise_conv2d_3	DepthwiseConv2D	depthwise	(3,3)	—	25 × 38 × 144
batch_normalization_10	BatchNormalization	—	—	—	25 × 38 × 144
re_lu_7	ReLU	—	—	—	25 × 38 × 144
conv2d_7	Conv2D	32 filtros	(1,1)	—	25 × 38 × 32
batch_normalization_11	BatchNormalization	—	—	—	25 × 38 × 32
conv2d_8	Conv2D	192 filtros	(1,1)	—	25 × 38 × 192
batch_normalization_12	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_8	ReLU	—	—	—	25 × 38 × 192
depthwise_conv2d_4	DepthwiseConv2D	depthwise	(3,3)	—	25 × 38 × 192
batch_normalization_13	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_9	ReLU	—	—	—	25 × 38 × 192
conv2d_9	Conv2D	32 filtros	(1,1)	—	25 × 38 × 32
batch_normalization_14	BatchNormalization	—	—	—	25 × 38 × 32
add_1	Add (conexión residual)	—	—	—	25 × 38 × 32
conv2d_10	Conv2D	192 filtros	(1,1)	—	25 × 38 × 192
batch_normalization_15	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_10	ReLU	—	—	—	25 × 38 × 192
depthwise_conv2d_5	DepthwiseConv2D	depthwise	(3,3)	—	25 × 38 × 192
batch_normalization_16	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_11	ReLU	—	—	—	25 × 38 × 192
conv2d_11	Conv2D	32 filtros	(1,1)	—	25 × 38 × 32
batch_normalization_17	BatchNormalization	—	—	—	25 × 38 × 32
add_2	Add (conexión residual)	—	—	—	25 × 38 × 32
conv2d_12	Conv2D	192 filtros	(1,1)	—	25 × 38 × 192
batch_normalization_18	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_12	ReLU	—	—	—	25 × 38 × 192
depthwise_conv2d_6	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 192
batch_normalization_19	BatchNormalization	—	—	—	13 × 19 × 192
re_lu_13	ReLU	—	—	—	13 × 19 × 192
conv2d_13	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_20	BatchNormalization	—	—	—	13 × 19 × 64
conv2d_14	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_21	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_14	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_7	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_22	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_15	ReLU	—	—	—	13 × 19 × 384
conv2d_15	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_23	BatchNormalization	—	—	—	13 × 19 × 64
add_3	Add (conexión residual)	—	—	—	13 × 19 × 64
conv2d_16	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_24	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_16	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_8	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_25	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_17	ReLU	—	—	—	13 × 19 × 384
conv2d_17	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_26	BatchNormalization	—	—	—	13 × 19 × 64
add_4	Add (conexión residual)	—	—	—	13 × 19 × 64
conv2d_18	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_27	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_18	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_9	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_28	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_19	ReLU	—	—	—	13 × 19 × 384
conv2d_19	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_29	BatchNormalization	—	—	—	13 × 19 × 64
add_5	Add (conexión residual)	—	—	—	13 × 19 × 64
conv2d_20	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_30	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_20	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_10	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_31	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_21	ReLU	—	—	—	13 × 19 × 384
conv2d_21	Conv2D	96 filtros	(1,1)	—	13 × 19 × 96
batch_normalization_32	BatchNormalization	—	—	—	13 × 19 × 96
conv2d_22	Conv2D	576 filtros	(1,1)	—	13 × 19 × 576
batch_normalization_33	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_22	ReLU	—	—	—	13 × 19 × 576
depthwise_conv2d_11	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 576
batch_normalization_34	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_23	ReLU	—	—	—	13 × 19 × 576
conv2d_23	Conv2D	96 filtros	(1,1)	—	13 × 19 × 96
batch_normalization_35	BatchNormalization	—	—	—	13 × 19 × 96
add_6	Add (conexión residual)	—	—	—	13 × 19 × 96
conv2d_24	Conv2D	576 filtros	(1,1)	—	13 × 19 × 576
batch_normalization_36	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_24	ReLU	—	—	—	13 × 19 × 576
depthwise_conv2d_12	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 576
batch_normalization_37	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_25	ReLU	—	—	—	13 × 19 × 576
conv2d_25	Conv2D	96 filtros	(1,1)	—	13 × 19 × 96
batch_normalization_38	BatchNormalization	—	—	—	13 × 19 × 96
add_7	Add (conexión residual)	—	—	—	13 × 19 × 96
conv2d_26	Conv2D	576 filtros	(1,1)	—	13 × 19 × 576
batch_normalization_39	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_26	ReLU	—	—	—	13 × 19 × 576
depthwise_conv2d_13	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 576
batch_normalization_40	BatchNormalization	—	—	—	7 × 10 × 576
re_lu_27	ReLU	—	—	—	7 × 10 × 576
conv2d_27	Conv2D	160 filtros	(1,1)	—	7 × 10 × 160
batch_normalization_41	BatchNormalization	—	—	—	7 × 10 × 160
conv2d_28	Conv2D	960 filtros	(1,1)	—	7 × 10 × 960
batch_normalization_42	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_28	ReLU	—	—	—	7 × 10 × 960
depthwise_conv2d_14	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 960
batch_normalization_43	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_29	ReLU	—	—	—	7 × 10 × 960
conv2d_29	Conv2D	160 filtros	(1,1)	—	7 × 10 × 160
batch_normalization_44	BatchNormalization	—	—	—	7 × 10 × 160
add_8	Add (conexión residual)	—	—	—	7 × 10 × 160
conv2d_30	Conv2D	960 filtros	(1,1)	—	7 × 10 × 960
batch_normalization_45	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_30	ReLU	—	—	—	7 × 10 × 960
depthwise_conv2d_15	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 960
batch_normalization_46	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_31	ReLU	—	—	—	7 × 10 × 960
conv2d_31	Conv2D	160 filtros	(1,1)	—	7 × 10 × 160
batch_normalization_47	BatchNormalization	—	—	—	7 × 10 × 160
add_9	Add (conexión residual)	—	—	—	7 × 10 × 160
conv2d_32	Conv2D	960 filtros	(1,1)	—	7 × 10 × 960
batch_normalization_48	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_32	ReLU	—	—	—	7 × 10 × 960
depthwise_conv2d_16	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 960
batch_normalization_49	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_33	ReLU	—	—	—	7 × 10 × 960
conv2d_33	Conv2D	320 filtros	(1,1)	—	7 × 10 × 320
batch_normalization_50	BatchNormalization	—	—	—	7 × 10 × 320
conv2d_34	Conv2D	1280 filtros	(1,1)	—	7 × 10 × 1280
batch_normalization_51	BatchNormalization	—	—	—	7 × 10 × 1280
re_lu_34	ReLU	—	—	—	7 × 10 × 1280
global_average_pooling2d	GlobalAveragePooling2D	—	—	—	1,280 elementos
Salida	Dense (Clasificador binario)	1 neuronas	—	Sigmoide	1 (Probabilidad de INE)

Veamos el sistema funcionando

Verificación de Identificaciones mediante Redes Neuronales

Resultados

Comportamiento en 30 épocas

Curva de entrenamiento (30 épocas)

Pérdida

Precisión

Entrenamiento

Validación

Final: 99.65% acc (99.2% val) · 0.0125 loss (0.0306 val)

Matriz de confusión

Pred. inePred. otro

Real ine

137

Real otro

2,820

Val Acc 99.4%
Diagonal (azul) = correcto. Fuera de diagonal (rojo) = error.

Espacio de features (capa densa, 512 dim)

PCAPC1 68.6% + PC2 23.1%

t-SNE

ine

otro

Distribución de pesos por capa

Bloque 1

μ=-0.005 · sp 0.23%

Bloque 2

μ=-0.011 · sp 0.00%

Bloque 3

μ=0.002 · sp 0.00%

Bloque 4

μ=-0.002 · sp 0.00%

conv2d_4

μ=0.003 · sp 0.03%

conv2d_5

μ=0.001 · sp 0.06%

conv2d_6

μ=0.002 · sp 0.06%

conv2d_7

μ=0.000 · sp 0.13%

conv2d_8

μ=0.000 · sp 0.05%

conv2d_9

μ=-0.002 · sp 0.08%

conv2d_10

μ=0.001 · sp 0.08%

conv2d_11

μ=-0.001 · sp 0.02%

conv2d_12

μ=-0.002 · sp 0.03%

conv2d_13

μ=0.000 · sp 0.09%

conv2d_14

μ=0.000 · sp 0.11%

conv2d_15

μ=-0.001 · sp 0.11%

conv2d_16

μ=0.000 · sp 0.09%

conv2d_17

μ=-0.001 · sp 0.09%

conv2d_18

μ=-0.000 · sp 0.07%

conv2d_19

μ=0.000 · sp 0.08%

conv2d_20

μ=0.000 · sp 0.08%

conv2d_21

μ=0.000 · sp 0.07%

conv2d_22

μ=-0.000 · sp 0.09%

conv2d_23

μ=0.001 · sp 0.10%

conv2d_24

μ=-0.000 · sp 0.11%

conv2d_25

μ=0.000 · sp 0.10%

conv2d_26

μ=0.000 · sp 0.08%

conv2d_27

μ=0.000 · sp 0.12%

conv2d_28

μ=-0.000 · sp 0.13%

conv2d_29

μ=-0.000 · sp 0.14%

conv2d_30

μ=0.000 · sp 0.14%

conv2d_31

μ=0.000 · sp 0.15%

conv2d_32

μ=-0.000 · sp 0.13%

conv2d_33

μ=0.000 · sp 0.13%

conv2d_34

μ=0.000 · sp 0.17%

Clasificación (512)

μ=0.017 · sp 0.31%

Desempeño

Impacto positivo y usabilidad

Diferenciación de credenciales similares: El modelo distingue las INE de credenciales de otras dependencias con paletas de colores parecidas.

Procesos con fotografía: Ideal para verificar identificaciones a partir de una fotografía.

Limitaciones identificadas

Inestabilidad por condiciones de iluminación y fondo: El modelo presenta baja certeza en la detección del objeto ante luz deficiente o sombras, al tiempo que genera falsos positivos con alta confianza frente a fondos homogéneos totalmente oscuros o brillantes.

Resolución de entrada: Al usar 200 × 300 píxeles, se restringe la legibilidad en al utilizar camaras.

Comparación con otros modelos

V1 (MFHA_CDO, el modelo propuesto) contra tres arquitecturas de referencia entrenadas sobre el mismo dataset — MobileNetV2, SqueezeNet y EfficientNet-B0.

Precisión de validación

V0.3 · EfficientNet-B0

Val. final: 99.8%

V1 · MFHA_CDO

V0.1 · MobileNetV2

V0.2 · SqueezeNet

V0.3 · EfficientNet-B0

Curva de entrenamiento

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

Matriz de confusión

Global 99.66%

2964/2974 aciertosBalanceada 99.20%INE 98.68% · 149/151

Global 99.43%

2957/2974 aciertosBalanceada 95.31%INE 90.73% · 137/151

Global 94.92%

2823/2974 aciertosBalanceada 50.00%INE 0.00% · 0/151

Global 99.73%

2966/2974 aciertosBalanceada 97.66%INE 95.36% · 144/151

Tamaño del modelo

20.47 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)

Tiempo de descarga del modelo

Referencia (V1)

~4.6× más rápido que V1

~14.3× más rápido que V1

~2.6× más rápido que V1

Tiempo de inferencia promedio

123 ms (base)

153 ms (1.2×)

144 ms (1.2×)

189 ms (1.5×)

Curva, matriz de confusión, tamaño y tiempo de descarga se leen de cada entrenamiento. El tiempo de inferencia es la excepción: es la mediana de 40 inferencias en CPU (promedio de dos pasadas).

Comparativa

Referencia ligera y eficiente frente al modelo propuesto V1

MobileNetV2

Bloques residuales invertidos

Cuellos de botella lineales

Convoluciones separables en profundidad

Salida sigmoide binaria

Arquitectura

Entrada

300 × 200 px RGB

Optimizer

RMSprop lr=1e-4

Loss

Binary Crossentropy

Épocas

Desliza la tabla para ver más →

Capa / Orden	Tipo de capa	Filtros / Neuronas	Kernel	Activación	Formato de salida
Entrada	Input Tensor	Canales RGB	—	—	200 × 300 × 3
conv2d	Conv2D	32 filtros	(3,3)	—	100 × 150 × 32
batch_normalization	BatchNormalization	—	—	—	100 × 150 × 32
re_lu	ReLU	—	—	—	100 × 150 × 32
depthwise_conv2d	DepthwiseConv2D	depthwise	(3,3)	—	100 × 150 × 32
batch_normalization_1	BatchNormalization	—	—	—	100 × 150 × 32
re_lu_1	ReLU	—	—	—	100 × 150 × 32
conv2d_1	Conv2D	16 filtros	(1,1)	—	100 × 150 × 16
batch_normalization_2	BatchNormalization	—	—	—	100 × 150 × 16
conv2d_2	Conv2D	96 filtros	(1,1)	—	100 × 150 × 96
batch_normalization_3	BatchNormalization	—	—	—	100 × 150 × 96
re_lu_2	ReLU	—	—	—	100 × 150 × 96
depthwise_conv2d_1	DepthwiseConv2D	depthwise	(3,3)	—	50 × 75 × 96
batch_normalization_4	BatchNormalization	—	—	—	50 × 75 × 96
re_lu_3	ReLU	—	—	—	50 × 75 × 96
conv2d_3	Conv2D	24 filtros	(1,1)	—	50 × 75 × 24
batch_normalization_5	BatchNormalization	—	—	—	50 × 75 × 24
conv2d_4	Conv2D	144 filtros	(1,1)	—	50 × 75 × 144
batch_normalization_6	BatchNormalization	—	—	—	50 × 75 × 144
re_lu_4	ReLU	—	—	—	50 × 75 × 144
depthwise_conv2d_2	DepthwiseConv2D	depthwise	(3,3)	—	50 × 75 × 144
batch_normalization_7	BatchNormalization	—	—	—	50 × 75 × 144
re_lu_5	ReLU	—	—	—	50 × 75 × 144
conv2d_5	Conv2D	24 filtros	(1,1)	—	50 × 75 × 24
batch_normalization_8	BatchNormalization	—	—	—	50 × 75 × 24
add	Add (conexión residual)	—	—	—	50 × 75 × 24
conv2d_6	Conv2D	144 filtros	(1,1)	—	50 × 75 × 144
batch_normalization_9	BatchNormalization	—	—	—	50 × 75 × 144
re_lu_6	ReLU	—	—	—	50 × 75 × 144
depthwise_conv2d_3	DepthwiseConv2D	depthwise	(3,3)	—	25 × 38 × 144
batch_normalization_10	BatchNormalization	—	—	—	25 × 38 × 144
re_lu_7	ReLU	—	—	—	25 × 38 × 144
conv2d_7	Conv2D	32 filtros	(1,1)	—	25 × 38 × 32
batch_normalization_11	BatchNormalization	—	—	—	25 × 38 × 32
conv2d_8	Conv2D	192 filtros	(1,1)	—	25 × 38 × 192
batch_normalization_12	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_8	ReLU	—	—	—	25 × 38 × 192
depthwise_conv2d_4	DepthwiseConv2D	depthwise	(3,3)	—	25 × 38 × 192
batch_normalization_13	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_9	ReLU	—	—	—	25 × 38 × 192
conv2d_9	Conv2D	32 filtros	(1,1)	—	25 × 38 × 32
batch_normalization_14	BatchNormalization	—	—	—	25 × 38 × 32
add_1	Add (conexión residual)	—	—	—	25 × 38 × 32
conv2d_10	Conv2D	192 filtros	(1,1)	—	25 × 38 × 192
batch_normalization_15	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_10	ReLU	—	—	—	25 × 38 × 192
depthwise_conv2d_5	DepthwiseConv2D	depthwise	(3,3)	—	25 × 38 × 192
batch_normalization_16	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_11	ReLU	—	—	—	25 × 38 × 192
conv2d_11	Conv2D	32 filtros	(1,1)	—	25 × 38 × 32
batch_normalization_17	BatchNormalization	—	—	—	25 × 38 × 32
add_2	Add (conexión residual)	—	—	—	25 × 38 × 32
conv2d_12	Conv2D	192 filtros	(1,1)	—	25 × 38 × 192
batch_normalization_18	BatchNormalization	—	—	—	25 × 38 × 192
re_lu_12	ReLU	—	—	—	25 × 38 × 192
depthwise_conv2d_6	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 192
batch_normalization_19	BatchNormalization	—	—	—	13 × 19 × 192
re_lu_13	ReLU	—	—	—	13 × 19 × 192
conv2d_13	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_20	BatchNormalization	—	—	—	13 × 19 × 64
conv2d_14	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_21	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_14	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_7	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_22	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_15	ReLU	—	—	—	13 × 19 × 384
conv2d_15	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_23	BatchNormalization	—	—	—	13 × 19 × 64
add_3	Add (conexión residual)	—	—	—	13 × 19 × 64
conv2d_16	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_24	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_16	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_8	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_25	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_17	ReLU	—	—	—	13 × 19 × 384
conv2d_17	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_26	BatchNormalization	—	—	—	13 × 19 × 64
add_4	Add (conexión residual)	—	—	—	13 × 19 × 64
conv2d_18	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_27	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_18	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_9	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_28	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_19	ReLU	—	—	—	13 × 19 × 384
conv2d_19	Conv2D	64 filtros	(1,1)	—	13 × 19 × 64
batch_normalization_29	BatchNormalization	—	—	—	13 × 19 × 64
add_5	Add (conexión residual)	—	—	—	13 × 19 × 64
conv2d_20	Conv2D	384 filtros	(1,1)	—	13 × 19 × 384
batch_normalization_30	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_20	ReLU	—	—	—	13 × 19 × 384
depthwise_conv2d_10	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 384
batch_normalization_31	BatchNormalization	—	—	—	13 × 19 × 384
re_lu_21	ReLU	—	—	—	13 × 19 × 384
conv2d_21	Conv2D	96 filtros	(1,1)	—	13 × 19 × 96
batch_normalization_32	BatchNormalization	—	—	—	13 × 19 × 96
conv2d_22	Conv2D	576 filtros	(1,1)	—	13 × 19 × 576
batch_normalization_33	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_22	ReLU	—	—	—	13 × 19 × 576
depthwise_conv2d_11	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 576
batch_normalization_34	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_23	ReLU	—	—	—	13 × 19 × 576
conv2d_23	Conv2D	96 filtros	(1,1)	—	13 × 19 × 96
batch_normalization_35	BatchNormalization	—	—	—	13 × 19 × 96
add_6	Add (conexión residual)	—	—	—	13 × 19 × 96
conv2d_24	Conv2D	576 filtros	(1,1)	—	13 × 19 × 576
batch_normalization_36	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_24	ReLU	—	—	—	13 × 19 × 576
depthwise_conv2d_12	DepthwiseConv2D	depthwise	(3,3)	—	13 × 19 × 576
batch_normalization_37	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_25	ReLU	—	—	—	13 × 19 × 576
conv2d_25	Conv2D	96 filtros	(1,1)	—	13 × 19 × 96
batch_normalization_38	BatchNormalization	—	—	—	13 × 19 × 96
add_7	Add (conexión residual)	—	—	—	13 × 19 × 96
conv2d_26	Conv2D	576 filtros	(1,1)	—	13 × 19 × 576
batch_normalization_39	BatchNormalization	—	—	—	13 × 19 × 576
re_lu_26	ReLU	—	—	—	13 × 19 × 576
depthwise_conv2d_13	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 576
batch_normalization_40	BatchNormalization	—	—	—	7 × 10 × 576
re_lu_27	ReLU	—	—	—	7 × 10 × 576
conv2d_27	Conv2D	160 filtros	(1,1)	—	7 × 10 × 160
batch_normalization_41	BatchNormalization	—	—	—	7 × 10 × 160
conv2d_28	Conv2D	960 filtros	(1,1)	—	7 × 10 × 960
batch_normalization_42	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_28	ReLU	—	—	—	7 × 10 × 960
depthwise_conv2d_14	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 960
batch_normalization_43	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_29	ReLU	—	—	—	7 × 10 × 960
conv2d_29	Conv2D	160 filtros	(1,1)	—	7 × 10 × 160
batch_normalization_44	BatchNormalization	—	—	—	7 × 10 × 160
add_8	Add (conexión residual)	—	—	—	7 × 10 × 160
conv2d_30	Conv2D	960 filtros	(1,1)	—	7 × 10 × 960
batch_normalization_45	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_30	ReLU	—	—	—	7 × 10 × 960
depthwise_conv2d_15	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 960
batch_normalization_46	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_31	ReLU	—	—	—	7 × 10 × 960
conv2d_31	Conv2D	160 filtros	(1,1)	—	7 × 10 × 160
batch_normalization_47	BatchNormalization	—	—	—	7 × 10 × 160
add_9	Add (conexión residual)	—	—	—	7 × 10 × 160
conv2d_32	Conv2D	960 filtros	(1,1)	—	7 × 10 × 960
batch_normalization_48	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_32	ReLU	—	—	—	7 × 10 × 960
depthwise_conv2d_16	DepthwiseConv2D	depthwise	(3,3)	—	7 × 10 × 960
batch_normalization_49	BatchNormalization	—	—	—	7 × 10 × 960
re_lu_33	ReLU	—	—	—	7 × 10 × 960
conv2d_33	Conv2D	320 filtros	(1,1)	—	7 × 10 × 320
batch_normalization_50	BatchNormalization	—	—	—	7 × 10 × 320
conv2d_34	Conv2D	1280 filtros	(1,1)	—	7 × 10 × 1280
batch_normalization_51	BatchNormalization	—	—	—	7 × 10 × 1280
re_lu_34	ReLU	—	—	—	7 × 10 × 1280
global_average_pooling2d	GlobalAveragePooling2D	—	—	—	1,280 elementos
Salida	Dense (Clasificador binario)	1 neuronas	—	Sigmoide	1 (Probabilidad de INE)

Resultados

Comportamiento en 30 épocas

Curva de entrenamiento (30 épocas)

Pérdida

Precisión

Entrenamiento

Validación

Final: 99.65% acc (99.2% val) · 0.0125 loss (0.0306 val)

Matriz de confusión

Pred. inePred. otro

Real ine

137

Real otro

2,820

Val Acc 99.4%
Diagonal (azul) = correcto. Fuera de diagonal (rojo) = error.

Espacio de features (capa densa, 512 dim)

PCAPC1 68.6% + PC2 23.1%

t-SNE

ine

otro

Distribución de pesos por capa

Bloque 1

μ=-0.005 · sp 0.23%

Bloque 2

μ=-0.011 · sp 0.00%

Bloque 3

μ=0.002 · sp 0.00%

Bloque 4

μ=-0.002 · sp 0.00%

conv2d_4

μ=0.003 · sp 0.03%

conv2d_5

μ=0.001 · sp 0.06%

conv2d_6

μ=0.002 · sp 0.06%

conv2d_7

μ=0.000 · sp 0.13%

conv2d_8

μ=0.000 · sp 0.05%

conv2d_9

μ=-0.002 · sp 0.08%

conv2d_10

μ=0.001 · sp 0.08%

conv2d_11

μ=-0.001 · sp 0.02%

conv2d_12

μ=-0.002 · sp 0.03%

conv2d_13

μ=0.000 · sp 0.09%

conv2d_14

μ=0.000 · sp 0.11%

conv2d_15

μ=-0.001 · sp 0.11%

conv2d_16

μ=0.000 · sp 0.09%

conv2d_17

μ=-0.001 · sp 0.09%

conv2d_18

μ=-0.000 · sp 0.07%

conv2d_19

μ=0.000 · sp 0.08%

conv2d_20

μ=0.000 · sp 0.08%

conv2d_21

μ=0.000 · sp 0.07%

conv2d_22

μ=-0.000 · sp 0.09%

conv2d_23

μ=0.001 · sp 0.10%

conv2d_24

μ=-0.000 · sp 0.11%

conv2d_25

μ=0.000 · sp 0.10%

conv2d_26

μ=0.000 · sp 0.08%

conv2d_27

μ=0.000 · sp 0.12%

conv2d_28

μ=-0.000 · sp 0.13%

conv2d_29

μ=-0.000 · sp 0.14%

conv2d_30

μ=0.000 · sp 0.14%

conv2d_31

μ=0.000 · sp 0.15%

conv2d_32

μ=-0.000 · sp 0.13%

conv2d_33

μ=0.000 · sp 0.13%

conv2d_34

μ=0.000 · sp 0.17%

Clasificación (512)

μ=0.017 · sp 0.31%

Desempeño

Impacto positivo y usabilidad

Diferenciación de credenciales similares: El modelo distingue las INE de credenciales de otras dependencias con paletas de colores parecidas.

Procesos con fotografía: Ideal para verificar identificaciones a partir de una fotografía.

Limitaciones identificadas

Resolución de entrada: Al usar 200 × 300 píxeles, se restringe la legibilidad en al utilizar camaras.

Verificación de Identificacionesmediante Redes Neuronales

Comparativa

MobileNetV2

Dataset

Arquitectura

Veamos el sistema funcionando

Resultados

Desempeño

Impacto positivo y usabilidad

Limitaciones identificadas

Comparación con otros modelos

Verificación de Identificacionesmediante Redes Neuronales

Comparativa

MobileNetV2

Dataset

Arquitectura

Veamos el sistema funcionando

Resultados

Desempeño

Impacto positivo y usabilidad

Limitaciones identificadas

Comparación con otros modelos

Verificación de Identificaciones
mediante Redes Neuronales

Verificación de Identificaciones
mediante Redes Neuronales