Table 14 Comparison of features of Classic-CNN and signaryNet

From: Leveraging digital acquisition and DPB based SignaryNet for localization and recognition of heritage inscription palaeography

Feature/Layer

Classic CNN

SignaryNet

Input shape

(64, 64, 1)

(64, 64, 1)

Conv layers

3 layers: (16, 32, 64 filters)

5 conv layers: (32 × 2, 64 × 2, 128)

Kernel size

(3, 3)

(3, 3)

Padding

Implicit (valid)

Explicitly padding = ‘same'

Batch normalization

No

After every Conv layer

Zero padding

No

After Batch Normalization

Dropout

One at 0.5 (after dense)

Multiple (0.25, 0.25, 0.4, 0.5)

MaxPooling2D

After each Conv

After blocks and final conv

Flatten layer output

2304

8192

Dense Layer (1)

512 units (ReLU)

256 units (ReLU)

Dense Layer (Output)

100 units (softmax)

100 units (softmax)

Activation Function

ReLU + softmax

ReLU + softmax

Optimizer

Adam (default)

Adam (lr = 0.001)

Loss function

sparse_categorical_crossentropy

categorical_crossentropy

Data Augmentation

None

Flip, Zoom, Rotation

Image rescaling

None

Rescaling(1./255)

Mixed precision

No

Yes (mixed_float16)

XLA (JIT) compilation

No

Enabled

Callbacks used

None

EarlyStopping + ReduceLROnPlateau + ModelCheckpoint

Epochs

40

50

Total parameters

1,253,756

2,263,236

Trainable parameters

1,253,756

2,262,596

Non-trainable parameters

0

640 (from BatchNorm layers)

Model size (float32)

~4.8 MB

~8.63 MB

Accuracy (expected)

Moderate

Higher (deeper, regularized)

Overfitting risk

High (no augmentation/dropout)

Lower (more dropout, batchnorm, augmentation)

Training speed

Fast (small model)

Optimized for speed

Generalization

Weaker

Stronger