Tsi-cnn-net: truly shift-invariant convolutional neural network for Indian sign language recognition system

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 2
• Main Authors: Ghorai, Anudyuti, Nandi, Utpal, Singh, Moirangthem Marjit, Changdar, Chiranjit, Paul, Bachchu, Chowdhuri, Partha, Pal, Pabitra
• Format: Journal Article
• Language: English
• Published: London Springer London 01.06.2025; Springer Nature B.V
• Subjects: Adaptive sampling; Artificial neural networks; Classification; Computer Science; Data augmentation; Datasets; Invariants; Original Article; Pattern Recognition; Recognition; Truly Shift-Invariant CNN; Indian sign language; Adaptive polyphase sampling; Classification; Consistency
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-025-01428-7

The majority of Indian sign language (ISL) recognition systems applied convolutional neural network (CNN) based deep neural networks. However, the output of CNN image classifiers may vary significantly with a little shift in input images. This shortcoming can be partially addressed by data augmentation, anti-aliasing, or blurring that do not work with different input patterns the network trained on and non-linear activation functions like ReLU, respectively. To deal with this short-coming, an ISL recognition approach has been presented using truly shift-invariant CNN. A sub-sampling strategy i.e. adaptive polyphase sampling (APS) has been applied to allow CNN truly shift-invariant. The proposed system is completely consistent to classification task. Furthermore, it offers significantly outstanding classification accuracy not only on Indian sign language datasets but also on datasets of other sign languages.