This work proposes Spike-HTR, a hybrid spiking recognizer that controls both the number of spiking steps and the number of width positions processed by the deep sequence mixer, and proposes a CTC-guided length reducer to reduce sequence computation.
Abstract
Handwritten Text Recognition (HTR) is computationally imbalanced in two ways: most image pixels are background, and many width-axis sequence positions are blank-dominated. This creates a mismatch for Spiking Neural Networks (SNNs): handwriting is observed as a static image, whereas spiking computation unfolds over timesteps. We propose Spike-HTR, a hybrid spiking recognizer that controls both the number of spiking steps and the number of width positions processed by the deep sequence mixer. To make a static image suitable for short-horizon spiking inference, InkCoder converts it into a coarse-to-fine input stream, where early steps cover broad stroke regions and later steps emphasize sharper stroke details. To reduce sequence computation, a CTC-guided length reducer keeps likely character or uncertain positions and compresses long blank-dominated stretches before deep mixing. With $T{=}2$, Spike-HTR trains only on target data, decodes without language models or lexicons, and reaches validation/test CERs of 3.5/5.4, 2.3/2.5, and 4.2/3.9 on IAM, LAM, and READ2016. Codes are available at https://github.com/QomolangmaH/SpikeHTR.
Lapis is proposed, a spiking attention mechanism that scores each token pair by the L1 distance between its query and key first-spike latency vectors under time-to-first-spike coding, and maps this distance to an affinity through a Laplacian kernel.
Kaiwen Tang, Jiaqi Zheng, Zixuan Zhu et al.· 0 citations
Handwritten Urdu text recognition remains a very challenging problem due to the cursive and complex nature of the Nastaliq script. Each writer has a unique style, and there is a severe lack of large, well-labeled datasets. These challenges make Urdu text recognition much harder than for many other languages. Here we propose a novel and effective approach for handwritten Urdu word recognition that addresses both data scarcity and feature complexity. We utilized conditional diffusion models to generate synthetic handwritten Urdu word images. These models produce realistic word samples covering a wide range of variations, such as stroke thickness, curves, and writing styles. The synthetic images augment the training data, helping the recognition model learn better in low-resource settings. For recognition, we employ a Neighborhood Attention Transformer (NAT) to extract robust visual features from both real and generated images. The visual embeddings are then fed into a pretrained BART language model configured as an encoder–decoder, which predicts the correct Urdu word. In this work, we first construct a baseline handwritten Urdu word dataset, addressing the lack of publicly available dataset for the text recognition. Building upon this foundational dataset, we leverage diffusion based generative approach to generate diverse and realistic synthetic handwritten samples, significantly alleviating the challenges of manual data collection. To the best of our knowledge, this is the first study to introduce both a real handwritten Urdu word dataset and a diffusion-generated synthetic dataset. In addition, we develop a unified word recognition model trained jointly on handwritten and printed Urdu word data, leading to improved recognition robustness and performance.
Wahid Hussain, Shahbaz Hassan, I. Hassan et al.· IEEE Access· 0 citations
Recent Handwritten Text Recognition (HTR) systems for Arabic and other right-to-left historical manuscripts have advanced exploration through convolutional neural network (CNN), recurrent, and Transformer-based models. However, degradation, weak diacritics, unstable baselines, and visually similar cursive letterforms still limit the recognition robustness. This paper presented an edge-aware line-level HTR framework that extends a CNN-Transformer baseline with a learnable edge-extraction channel and Squeeze-and-Excitation (SE) channel attention. The edge module emphasized stroke boundaries and character contours, while SE attention recalibrated feature responses to suppress background artifacts and preserve informative ink patterns. The resulting sequence was modeled by a Transformer encoder and trained using Connectionist Temporal Classification (CTC) with an auxiliary decoder cross-entropy loss. The experiments on the Kalima Arabic manuscript line-image dataset, using Books 1-8 with 86 pages and 1,759 annotated text lines, reduced character error rate from 6.40% to 4.10% and word error rate from 27.43% to 20.43%. These results show that combining learnable structural cues with channel-wise attention has improved robustness for degradation-prone historical manuscript collections.
Bilal Abdulrahman, Farhan Mohamed· Journal of Human Centered Te...· 0 citations
Handwritten Text Recognition (HTR) has experienced remarkable progress with the development of deep learning techniques. However, most existing studies focus on high-resource languages for which large annotated datasets are readily available. In contrast, low-resource languages remain largely underrepresented in handwriting recognition research due to the scarcity of handwritten corpora, linguistic resources, and benchmark datasets. This paper presents a handwritten word recognition framework for Kirundi, a low-resource Bantu language spoken primarily in Burundi. The proposed system employs a Convolutional Recurrent Neural Network (CRNN) combined with Connectionist Temporal Classification (CTC) for end-to-end sequence recognition without explicit character segmentation.
To address severe data scarcity, a small handwritten Kirundi dataset consisting of manually collected word samples was constructed and annotated. Data augmentation techniques, including rotation, translation, Gaussian noise, Gaussian blur, and elastic distortion, were applied to increase sample diversity. In addition, synthetic handwritten-style data were generated to further expand the training set. Three experimental configurations were investigated: real handwritten data only, real data with augmentation, and real data combined with augmentation and synthetic handwritten-style images.
Experimental results demonstrate that synthetic data generation improved recognition performance and reduced Character Error Rate (CER) from 0.8354 to 0.7560, corresponding to an approximate relative improvement of 9.5%. Although exact word-level recognition remained difficult because of the extremely limited dataset size, the proposed framework successfully learned meaningful sequential patterns and produced increasingly structured Kirundi-like predictions. The study establishes an initial benchmark for Kirundi handwritten word recognition and highlights the potential of synthetic data generation for low-resource handwriting recognition tasks.
Niyifasha Patrick· International journal of re...· 0 citations
PTQ4SNN is proposed, a membrane-aware post-training quantization framework that jointly quantizes weights and recurrent membrane states using only a small calibration set and effectively preserves model accuracy under W4 quantization and approximately 4-bit membrane precision.
Hui Xie, Tong Shi, Haotong Qin et al.· 0 citations
Implantable brain-computer interfaces require on-node spike sorting to reduce telemetry bandwidth and power while maintaining reliable neural decoding. This paper presents a hardware-oriented deep binarized neural network (DBNN) spike-sorting system with two binarized hidden layers with 256 neurons and a fixed-point output layer to enable multiplier-free inference dominated by sign-controlled accumulation and bit-wise logic. The proposed classifier operates on compact 16-sample spike waveforms to reduce the implementation cost (16-256-256-3) and achieves a median classification accuracy of 98.7% on both synthetic and in-vivo datasets. An FPGA prototype on a Cyclone V device operates at 50 MHz and requires 528 cycles per spike, corresponding to a 0.01 ms compute latency, while consuming 828 ALMs and 1023 registers with zero DSP blocks. For ASIC feasibility, the DBNN is implemented using FreePDK45-based flow; synthesis in Synopsys Design Compiler indicates an estimated silicon area of 0.014 mm2 and an operating power of 122 nW at 20 kHz under a 1.1 V supply. These results demonstrate that the proposed DBNN spike sorter offers a favorable trade-off between accuracy and implementation cost, supporting low-power, implantable neural interfaces. Overall, the proposed DBNN spike sorter achieves high accuracy (98.7%) with extremely low hardware cost (0.014 mm2, 122 nW at 20 kHz) and multiplier-free operation, making it suitable for low-power, implantable neural interfaces. This paper introduces the first DBNN designed for real-time neural spike sorting, striking an excellent balance between input data size and network complexity.