<![CDATA[The approach presents a multi braille character (MBC) recognition system for Indonesian syllablesdesigned to address real-world imaging variations. The proposed framework formulates 105-class visual classification task, where each class represents a two-character Braille unit. This design aims to preserve inter-character spatial relationships and reduce error propagation commonly found in single-character segmentation approaches. A carefully constructed dataset undergoes spatial pre-processing stages, including rotation normalization, grid assignment, and multicell cropping, resulting in uniform 89×89 pixel image patches that ensure geometric consistency across samples. To enhance model generalization under varying illumination conditions, single-dimension photometric augmentation is applied exclusively during training, including brightness (±25%), exposure (±20%), saturation (±40%), and hue (±30%). ResNet-101 is adopted as the backbone architecture based on prior comparative studies conducted on the same dataset, demonstrating its effectiveness in capturing fine-grained Braille dot shadow patterns. The network is trained for 300 epochs with a batch size of 32 under consistent experimental settings, and performance is evaluated using a confusion-matrix-based framework with overall accuracy as the primary metric. Experimental results indicate that moderate photometric reductions significantly improve recognition performance by preserving critical micro-contrast cues. In particular, an exposure reduction of −20% achieves the best balance between accuracy (86.13%) and training efficiency (14.12 minutes), outperforming the non-augmented baseline (74.37%, 22.10 minutes). A hue reduction of −30% further improves robustness to ambient color variations, while aggressive positive adjustments degrade performance due to structural distortion. These findings confirm the effectiveness of the proposed MBC framework for practical Braille recognition in real-world environments.]]>
