Recognizing Manchu words can be challenging due to their complex character variations, subtle differences between similar characters, and homographic polysemy. Most studies rely on character segmentation techniques for character recognition or use convolutional neural networks (CNNs) to encode word images for word recognition. However, these methods can lead to segmentation errors or a loss of semantic information, which reduces the accuracy of word recognition. To address the limitations in the long-range dependency modeling of CNNs and enhance semantic coherence, we propose a hybrid architecture to fuse the spatial features of original images and spectral features. Specifically, we first leverage the Short-Time Fourier Transform (STFT) to preprocess the raw input images and thereby obtain their multi-view spectral features. Then, we leverage a primary CNN block and a pair of symmetric CNN blocks to construct a symmetric spectral enhancement module, which is used to encode the raw input features and the multi-view spectral features. Subsequently, we design a feature fusion module via Swin Transformer to fuse multi-view spectral embedding and thereby concat it with the raw input embedding. Finally, we leverage a Transformer decoder to obtain the target output. We conducted extensive experiments on Manchu words benchmark datasets to evaluate the effectiveness of our proposed framework. The experimental results demonstrated that our framework performs robustly in word recognition tasks and exhibits excellent generalization capabilities. Additionally, our model outperformed other baseline methods in multiple writing-style font-recognition tasks.
Optical Character Recognition (OCR) for Manchu is hindered by large intra-domain variation caused by diverse fonts. Consequently, traditional end-to-end models perform poorly in zero-shot cross-domain scenarios. To address this bottleneck, we propose a two-stage decoupled paradigm, “Normalize-then-Recognize,” and construct the Target-Masked Adversarial Reconstruction Network (TMAR-Net) as a modular visual pre-filter. To address the continuous cursive axis and severe foreground–background imbalance of Manchu script, TMAR-Net utilizes a paired ConvNeXt generator equipped with bottleneck self-attention and introduces a Structure-Aware Hard-Mining Target-Masked Loss (SA-HMTM Loss). Through the joint application of pixel-level masking, instance-level hard mining, and feature-level structural anchoring, this approach reduces broken strokes, long-tail sample collapse, and topological hallucinations during reconstruction. Evaluated on a dataset containing 20,000 words, system-level validation shows that, after lightweight domain adaptation, the end-to-end recognition accuracies of single-font and multi-font baseline models rise to averages of 92.34% and 92.56%, respectively, peaking at 95.52%. This yields an average absolute improvement of over 90 percentage points compared with the direct single-font baseline, closely approaching the 99.81% recognition rate typically achieved under in-domain conditions. This work provides a robust, system-level solution for digitizing high-variance, low-resource minority multi-font documents.