uTHCD-C: A Corruption-Robustness Benchmark for Tamil Handwritten Character Recognition
Authors: Preethi S, Meeradevi T, Arun Prakash A
Organizations: Dept. of Electronics and Communication Engineering, Kongu Engineering College, Erode, Tamil Nadu, India · AI4Bharat, Indian Institute of Technology Madras, Chennai, Tamil Nadu, India
Evaluation of Tamil handwritten character recognition (HCR) has largely focused on clean test images, and a standardized multi-corruption robustness benchmark for isolated Tamil HCR has not been established. We introduce uTHCD-C, a corruption-robustness benchmark based on the unconstrained Tamil Handwritten Character Database (uTHCD). The benchmark defines ten corruption types, including a handwriting-specific stroke-thinning corruption, at five severity levels. We evaluate nine convolutional and transformer-based architectures using five-fold stratified cross-validation and writer-independent testing. Robustness is quantified using mean Corruption Error (mCE), normalized against a VGG16-BN reference. We additionally compare parameter count, computational complexity, and inference throughput to characterize the computational requirements of the evaluated architectures. The results establish comparative robustness baselines for Tamil HCR, revealing substantial differences across architectures: ConvNeXt and Swin achieve the lowest mCE values (0.75 and 0.83, respectively), whereas ShuffleNet and MnasNet exhibit higher values (1.55 and 1.85, respectively). The accompanying efficiency measurements characterize the trade-offs between corruption robustness and computational requirements. By providing standardized corruption configurations, evaluation protocols, and comparative results, uTHCD-C establishes a reproducible basis for future research on robust and computationally efficient Tamil HCR.
Figures & tables
Study
Script and task
Degradation evaluation
Role relative to uTHCD-C
Shaffi and Hajamohideen (2021)
Tamil, isolated HCR
Clean images only
Provides the uTHCD dataset that uTHCD-C extends
Kavitha et al. (2022)
Tamil, isolated HCR
Clean images only
Recognition baseline on uTHCD without corruption evaluation
Kavitha and Srimathi (2022)
Tamil, isolated HCR
Clean images only
CNN recognition baseline without corruption evaluation
Karki et al. (2018)
Bangla, characters and numerals
Pixel-level noise (reconstruction and classification)
Addresses noise handling for a different script, not a benchmark
Chaudhury et al. (2022)
Bangla, degraded-document OCR
Degraded scanned documents
Document-level OCR, not an architecture-level robustness benchmark
This work (uTHCD-C)
Tamil, isolated HCR
Ten corruption types × five severity levels
Standardized multi-corruption benchmark comparing nine architectures via mCE
Table 1: Comparison of relevant prior work and the proposed uTHCD-C benchmark.
Characteristic
Value
Dataset
uTHCD
Script
Tamil
Number of samples
90,950
Number of classes
156
Number of writers
753
Native image resolution
64×64×3
Table 2: Summary of the uTHCD dataset.
Model
TorchVision variant
VGG16-BN
vgg16_bn
GoogLeNet
googlenet
EfficientNet-B0
efficientnet_b0
RegNetX-400MF
regnet_x_400mf
MnasNet-0.5
mnasnet0_5
ShuffleNet V2-0.5
shufflenet_v2_x0_5
Table 3: Evaluated architectures and their exact torchvision variants.
Corruption
Parameter
s1
s2
s3
s4
s5
Gaussian noise
std. dev. σ
10
20
30
40
60
Shot noise
Poisson scale
60
40
25
15
8
Impulse noise
corruption prob.
0.03
0.06
0.12
0.18
0.25
Gaussian blur
radius
0.7
1.5
2.5
3.5
5.0
Defocus blur
radius
1
2
3
4
6
Stroke thinning
re-dilation kernel
9
7
3
1
1
Table 4: Corruption types and parameter settings for the five nominal severity levels. Parameter effects depend on the corruption operator: larger Gaussian-noise, impulse-noise, blur, and elastic-deformation parameters produce stronger corruption, whereas smaller shot-noise scales, contrast factors, and stroke-thinning re-dilation kernels produce stronger degradation. Pixelation uses the listed downsampling factor before restoring the image to its original resolution. For scale corruption, the listed intermediate target sizes are followed by restoration to the native 64×64 resolution; target sizes of 128 and 64 pixels do not constitute downsampling from the native resolution. Stroke-thinning levels 4 and 5 use the same kernel size of 1 and therefore produce identical transformations.
Figure 1: The ten corruption types in uTHCD-C applied to a single representative Tamil character at a fixed severity level (severity 3). Stroke thinning is the handwriting-specific operator; the remaining nine corruptions are adapted from established corruption benchmarks.
Model
Clean accuracy (%)
ConvNeXt
97.82 ± 0.13
EfficientNet
97.77 ± 0.11
RegNet
97.73 ± 0.19
Swin
97.57 ± 0.14
ShuffleNet
97.27 ± 0.14
MnasNet
97.11 ± 0.24
Table 5: Clean-data recognition accuracy under 5-fold stratified cross-validation (mean ± standard deviation over five folds, sorted by accuracy). The reference model is shown in bold.
Model
Clean acc. (%)
Params (M)
GMACs
Throughput (img/s)
ShuffleNet
97.27
0.50
0.04
30,591
SqueezeNet
95.74
0.81
0.75
7,240
MnasNet
97.11
1.14
0.10
12,750
EfficientNet
97.77
4.21
0.38
5,110
RegNet
97.73
5.16
0.41
10,242
GoogLeNet
96.33
5.70
1.50
6,479
Table 6: Resource characteristics of the evaluated architectures: parameter count (classifier resized to 156 classes), computational complexity (GMACs for one 224×224 forward pass), and inference throughput (images/s, measured on an NVIDIA A100 in FP32 at 224×224 ; higher is faster). Clean accuracy (5-fold mean) is repeated for reference. Sorted by parameter count. The reference model is shown in bold.
Model
mCE
ConvNeXt
0.75 ± 0.03
Swin
0.83 ± 0.07
EfficientNet
0.91 ± 0.16
GoogLeNet
0.98 ± 0.16
VGG16-BN
1.00 ± 0.00
SqueezeNet
1.10 ± 0.15
Table 7: Aggregate corruption robustness under 5-fold stratified cross-validation, measured by mCE (VGG16-BN reference, lower is better; mean ± standard deviation over five folds, sorted by mCE). The reference model is shown in bold.
Model
Clean (%)
Gaussian σ=30 (%)
ConvNeXt
97.82
97.81
Swin
97.57
97.57
VGG16-BN
97.00
96.44
GoogLeNet
96.33
94.71
SqueezeNet
95.74
94.91
EfficientNet
97.77
84.69
Table 8: Clean-data accuracy and accuracy under Gaussian noise ( σ=30 ), 5-fold mean (%). Models with similar clean accuracy differ sharply under additive noise.
Figure 2: Per-corruption mCE by architecture (5-fold mean, ratio-of-sums definition; VGG16-BN =1.0 , lower is better). Rows are ordered by aggregate mCE and the rightmost column repeats the aggregate value. The largest cross-architecture differences occur under Gaussian and shot noise, with additional variation under several other corruption types. The color scale is capped at 3.0 ; cells exceeding this value are shown saturated, and the annotations report the exact mCE.
Model
Writer-dependent mCE
Writer-independent mCE
ConvNeXt
0.81
0.81 ± 0.03
Swin
0.85
0.92 ± 0.05
VGG16-BN
1.00
1.00 ± 0.00
GoogLeNet
0.97
1.13 ± 0.24
EfficientNet
0.88
1.14 ± 0.11
SqueezeNet
1.13
1.32 ± 0.09
Table 9: Writer-independent corruption robustness (mCE, VGG16-BN reference, lower is better). “Writer-dependent” is the single fixed 70:30 split; “writer-independent” is the mean ± standard deviation over three writer-disjoint partitions (seeds 1 , 42 , 123 ). Sorted by writer-independent mCE; the reference model is shown in bold.
Figure 3: Pareto frontier of inference throughput and corruption robustness. Each point represents one evaluated architecture. Higher throughput and lower mCE are preferable. Pareto-optimal models are those for which no other evaluated architecture is simultaneously faster and more robust. The frontier is defined over these two criteria only.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Model
Gauss.
Shot
Impulse
G.blur
Defoc.
Thin
Elastic
Pixel.
Contr.
Scale
VGG16-BN
86.1
66.0
11.1
22.9
19.1
95.8
94.0
70.5
77.2
57.2
GoogLeNet
90.0
76.9
14.5
43.1
35.3
94.0
91.2
62.2
76.8
71.3
EfficientNet
71.5
58.2
13.4
58.5
49.1
96.5
94.4
70.1
96.8
77.0
RegNet
59.3
26.8
7.9
48.1
38.7
96.5
93.5
67.1
79.4
74.3
MnasNet
0.8
0.7
10.3
41.2
32.3
94.9
93.2
64.9
80.9
66.9
ShuffleNet
25.7
4.6
9.4
55.6
46.6
96.0
92.7
69.3
81.1
75.3
Appendix
Table 10: Accuracy (%) under each corruption, averaged over the five severity levels (5-fold mean); higher is better. The noise columns (Gauss., Shot, Impulse) carry the largest differences across architectures. Full per-severity tables are available in the released benchmark.
We built a compact convolutional network (1.11 M parameters) for 46-class DHCD Devanagari recognition and reached 99.73%, the highest reported at 15.6x smaller than prior state-of-the-art. We have effectively reached the saturation point: every model tested, large teacher ensembles included, hits the same 11-error intrinsic floor. No configuration achieves a statistically clear win under exact McNemar tests with Wilson confidence intervals. Even without knowledge distillation, our student matches the nearest large-model baseline (17.32 M parameters; McNemar p=0.345). Outside of DHCD, zero-shot on CMATERdb digits gives 76.6% and fine-tuning reaches 97.8%; corruption robustness is also far better than large baselines (mean corruption accuracy 75.7% vs. 38.7%). All artifacts are at https://github.com/Ampixa/barnamala.
Fine-tuning transformer-based handwritten text recognition (HTR) models on medieval manuscripts is challenging because these models are pre-trained on modern text and must adapt to a very different visual domain. This paper studies how three controllable fine-tuning choices (contrast normalization, data augmentation, and layer freezing) affect recognition accuracy when adapting TrOCR to small historical datasets. We run controlled experiments on a 13th-century Italian manuscript (I-CT 91 "Cortonese") and replicate the same experimental grid on the public READ-16 benchmark as robustness evidence. On Cortonese, our best configuration achieves 8.03% character error rate (CER). Statistical comparisons across 13 configurations show that freezing up to three encoder layers or six decoder layers does not significantly harm accuracy, while deeper freezing becomes progressively detrimental. Removing contrast normalization (CLAHE) yields 7.84% CER, comparable to a domain-specialized baseline, suggesting strong optimization can reduce reliance on image preprocessing. Cross-dataset validation on READ-16 shows that decoder freezing thresholds transfer more robustly than encoder thresholds, and combined freezing strategies require dataset-specific re-validation. Finally, we use Grad-CAM gradient attributions and decoder cross-attention maps to diagnose error patterns and failure modes revealed by the ablations. Source code is available at https://github.com/LaudareProject/TrOCR-analysis
Non-Latin handwritten character recognition (HCR) remains understudied. Dominant methods consider it as generic image classification, which uses model scale to implicitly learn stroke structure. Structural-prior efficiency---the principle that explicitly encoding script-geometric regularities as architectural inductive biases can be both more accurate and require fewer parameters. We introduce GraphemeNet, a unified multi-script architecture, governed by two orthogonal binary axes. Axis 1 operationalises stroke-level geometric regularity via Persistent Scaffold Injection (PSI): a script-specific asymmetric convolution injects a stroke scaffold as a weighted residual at every encoder stage, continuously anchoring learned features to script geometry---distinct from skip connections, auxiliary losses, or attention reweighting. Axis 2 selects between global average pooling with gated fusion and cross-scale attention with a Stroke Topology Module (STM), depending on whether glyph discrimination requires spatial relational reasoning. A Linear Capsule Routing (LCR) with O(n) routing is shared universally. On fourteen benchmarks across eight writing systems, the architecture generalises with only scaffold and decoder topology varying per script, consistently challenging, outperforming published baselines, and establishing structural-prior efficiency as a broadly applicable principle for multi-script HCR.
Ranjit Raut, Aarav Subedi, Ashim Shrestha
Department of Artificial Intelligence Kathmandu University Dhulikhel, Nepal