<![CDATA[Hearing and speech disabilities isolate millions from everyday social interaction; automated sign language interpretation offers a practical avenue for bridging this divide. This paper presents a Convolutional Neural Network that classifies 24 static hand postures constituting the American Sign Language (ASL) alphabet. Three distinguishing design decisions shape the proposed system: per-block batch normalization that stabilizes activation statistics at every spatial resolution stage, stochastic dropout applied at two architectural depths, and a chirality-preserving augmentation pipeline that deliberately omits axis-flip transforms because mirrored ASL handshapes encode different letters. Trained on 27,455 grayscale images drawn from the Sign Language MNIST repository and evaluated on a held-out partition of 7,172 samples, the network achieves 99.85% training accuracy and 100.00% test accuracy after 20 epochs with Adam optimization. All 24 gesture categories attain precision, recall, and F1-score of 1.00. At 263,749 trainable weights totalling roughly 1 MB, the architecture proves viable for deployment on edge and mobile hardware without pre-trained backbone dependencies.]]>
Copyrights © 2026
