Family Balancing and Hierarchical CNNs for Arabic Handwritten Symbols: A Controlled Five-Seed Evaluation

Abstract A joint Arabic handwriting recognizer must distinguish digits and letters despite unequal representationin its training data. We test whether family-balanced learning and hierarchical classification improvethe less accurately recognized family. To examine this question, six convolutional neural network (CNN)configurations used the same four-block feature-extractor architecture and were trained with fivematched random seeds on a 38-class task assembled from MADBase digits and the Arabic HandwrittenCharacters Dataset. Comparisons separated equal-family loss weighting, hierarchical prediction, smoothemphasis on the higher-loss family, and parameter count. Exact-image-separated partitions contained62,403 training, 11,014 validation, and 13,360 test images. Each selected model was evaluated on cleanimages and six fixed corruptions, producing 210 evaluation records. Across these configurations, overallclean-image accuracy ranged from approximately 98.2% to 98.4%. The lower of the digit and letterclass-averaged F1 scores was 0.9637 for the flat baseline and 0.9658 for its family-balanced counterpart.Their paired difference was +0.0021 F1, with a descriptive five-seed interval of [−0.0076,0.0118].Hierarchical prediction produced mixed clean-image effects. Smooth worst-family training reducedmean corruption retention by 0.0317 F1 relative to balanced hierarchical training. Additive noise andmidpoint-centered contrast changes caused substantially larger performance losses. Taken together,the controlled comparisons identify small, inconsistent gains from family-aware training and outputorganization, alongside pronounced sensitivity to the specified image transformations. Family-specificevaluation reveals these differences and directs subsequent work toward understanding the observedcorruption response.