HMAFNet: A Hierarchical Multi-Scale Attention Fusion Network for Offline Handwritten Odia Compound Character Recognition

Abstract

Handwriting based offline Odia compound character recognition could be challenging mainly due to the complex structure of the conjunct characters along with having significant variations in handwriting style. Moreover, touching characters, discontinuous strokes, degradation in documents, and so forth make the problem even harder. Despite promising performance of the recently introduced convolution-based and transformer-based architectures, there is still the lack of their understanding of capturing multi scale information needed to recognize these complex Odia compound characters. In order to address these issues, HMAFNet – Hierarchical Multi-Scale Attention Fusion Network, which is intended for offline handwritten Odia compound character recognition, is introduced in this paper. In particular, the architecture of the suggested system is comprised of two main parts: first, efficient preprocessing and segmentation algorithm; second, three complementary modules. To start with, the first module Hierarchical Multi-Scale Feature Extraction (HMFE) is proposed to learn discriminative multi resolution representations. Secondly, Multi-Scale Attention Fusion (MAF) module adaptively emphasizes the most informative channel and spatial information. Thirdly, Hierarchical Feature Aggregation (HFA) module is responsible for integration of hierarchical semantic information. In order to test the proposed model, it is assessed using the public ODIA handwritten compound characters data set named OHCD_GIETV2, and the comparison is made with a few state-of-the-art deep learning architectures such as ResNet-50, DenseNet-121, EfficientNet-B0, ConvNeXt-Tiny, Swin Transformer-T, MaxViT-T, FasterViT-0, and ConvNeXt V2-Tiny. The experimental results suggest that HMAFNet achieves the recognition accuracy of 89.26%, 87.23% precision, 88.19% recall, and F1-Score of 87.71%. Additionally, Character Error Rate (CER) reduces to 0.96%. In addition to that, the number of parameters required by the network is only 18.72M and the inference time is just 7.4 ms per image, which suggests a decent balance between recognition performance and computational efficiency. All of these results tend to verify that HMAFNet is a reliable, efficient choice for offline handwriting Odia compound characters recognition, and in particular provides huge potential for the application in digitization of documents and archiving systems, specifically for low-resource Indic scripts.