Optimizing MNIST Digit Recognition for Edge Devices: A Multi-dimensional Evaluation Approach

Handwritten Digit Recognition Based on Convolutional Neural Network and Its Application in Overlapping Image Processing

Optimizing 1001-Class Handwritten Digit Sequence Recognition for the Mexican Electoral Process Using Asymmetric ResNet-18 and CBAM

The automated digitization of handwritten electoral results is critical for ensuring transparency and speed in democratic processes. While recurrent sequence-to-sequence models (e.g., CRNN+CTC) achieve high accuracy, they inherently violate the strict latency constraints of high-throughput administrative environments. Conversely, standard lightweight CNNs exhibit suboptimal performance on the long-tail distribution of high-cardinality scenarios. To bridge this gap, this study reformulates sequence recognition into a latency-bound classification task. We propose a specialized Handwritten Digit Sequence Recognition (HDSR) framework for the Mexican Preliminary Election Results Program (PREP) based on a modified ResNet-18 architecture. The methodology introduces an asymmetric stride designed explicitly to preserve the 1:3 horizontal feature resolution of electoral tally sheets, integrating a lightweight Convolutional Block Attention Module (CBAM) in deep stages to refine classification across 1001 possible sequences. Leveraging a megadiverse dataset of 3.77 million real-world images, the model was trained using AdamW and label smoothing to mitigate human-induced label noise. Results demonstrate a global accuracy of 97.82% and a significant improvement in Macro-Precision (0.8878) for rare sequences. With an inference latency of 9.1 ms on standard CPU hardware, the proposed solution offers a scalable, high-confidence alternative that prioritizes spatial preservation and fail-controlled deployment.

A Dynamic Stochastic Computing-Based Approach to Edge Inference for Convolutional Neural Networks

Deep neural networks achieve state-of-the-art performance in many vision tasks but require substantial computational resources and energy, making their deployment on edge devices and embedded platforms challenging. Several approaches have been proposed to address this challenge, including fixed-point arithmetic, stochastic computing (SC), and dynamic stochastic computing (DSC). This work presents a comparative analysis of these approaches using a convolutional neural network (CNN) for MNIST digit classification. The study compares floating-point inference, fixed-point quantized inference, stochastic computing using bitstream representations, and dynamic stochastic computing using accumulation with stochastic integrators. Performance is evaluated in terms of classification accuracy, cross-entropy loss, runtime latency, and throughput across different bit-widths and stochastic bitstream lengths. The results demonstrate the trade-offs among numerical precision, computational latency, and inference accuracy. Fixed-point implementations provide high throughput with minimal accuracy degradation at moderate bit-widths, whereas stochastic computing reduces arithmetic complexity but requires longer bitstreams to achieve comparable accuracy. Dynamic stochastic computing exhibits improved convergence characteristics compared with conventional stochastic computing. These results indicate that dynamic stochastic computing can serve as an efficient alternative for the design and implementation of neural-network accelerators.

Barnamala: Parameter-Efficient Handwritten Devanagari Recognition at Benchmark Saturation

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.

A Comparative Study of Deep Learning Architectures for Image Classification

Image classification is a major area in computer vision, driven by rapid advances in deep learning. Over the last decade, convolutional neural networks (CNNs) and their variants have achieved high performance in applications such as medical diagnosis, autonomous driving, industrial inspection, remote sensing, and biometrics. However, choosing the right model remains challenging due to trade-offs between accuracy, computational cost, efficiency, and robustness. This paper presents a comparative study of different deep learning architectures, including classical CNNs, deep hierarchical models, residual and dense networks, and compound-scaled architectures. Using a common evaluation framework and standard datasets, the study analyzes performance based on key design factors such as depth, width, receptive field, skip connections, and normalization. Theoretical concepts like convolution operations, residual learning, and optimization are also discussed.The results show that deeper networks provide better representation, while residual connections and compound scaling improve training stability and efficiency. Lightweight models perform well in resource-limited and real-time environments. Overall, the study offers practical guidance for selecting suitable architectures and highlights future research areas such as neural architecture search, self-supervised learning, and efficient model deployment.

Edge-native static American Sign Language recognition using a lightweight geometric multilayer perceptron

The global imperative for inclusive communication technologies necessitates the development of highly accurate, computationally efficient Sign Language Recognition (SLR) systems capable of deployment on resource-constrained edge devices. Current state-of-the-art approaches overwhelmingly rely on deep Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), which possess memory footprints and latency profiles incompatible with ultra-low-power microcontrollers. This report presents the architectural formulation, mathematical engineering, and empirical validation of an edge-native American Sign Language (ASL) recognition framework. By shifting the computational burden from brute-force pixel processing to explicit geometric feature engineering, the system utilizes the MediaPipe hand landmark regressor to extract a 21-point 3D topological skeleton. This spatial data is mathematically compressed into a dense, 30-dimensional tensor of scale-invariant geometric features, including intrinsic joint angles, normalized inter-phalangeal distances, and orthogonal projections. Classification is executed via a highly regularized, four-layer Multi-Layer Perceptron (MLP) with a memory footprint of merely 299 kilobytes. Evaluated against the Kaggle ASL Alphabet dataset, the 20-feature optimally pruned model achieved a 99.21% global F1-score outperforming the baseline model with 98.66% accuracy. The findings demonstrate a paradigm shift in TinyML assistive technology, proving that explicit mathematical sparsity outperforms high-dimensional deep learning for static spatial taxonomies.