(IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 14, No. 5, 2023 QMX-BdSL49: An Efficient Recognition Approach for Bengali Sign Language with Quantize Modified Xception Nasima Begum*, Saqib Sizan Khan, Rashik Rahman, Ashraful Haque, Nipa Khatun, Nusrat Jahan, Tanjina Helaly Department of Computer Science and Engineering, University of Asia Pacific, Dhaka, Bangladesh Abstract—Sign language is developed to bridge the com- munication gap between individuals with and without hearing impairment or speech difficulties. Individuals with hearing and speech impairment typically rely on hand signs as a means of expressing themselves. However, people, in general, may not have sufficient knowledge of sign language, thus a sign language recognition system on an embedded device is most needed. Literature related to such systems on embedded devices is scarce as these recognition tasks are very complex and computationally expensive. The limited resources of embedded devices cannot execute complex algorithms like Convolutional Neural Network (CNN) properly. Therefore, in this paper, we propose a novel deep learning architecture based on default Xception architec- ture, named Quantized Modified Xception (QMX) to reduce the model’s size and enhance the computational speed without compromising model accuracy. Moreover, the proposed QMX model is highly optimized due to the weight compression of model quantization. As a result, the footprint of the proposed QMX model is 11 times smaller than the Modified Xception (MX) model. To train the model, BDSL 49 dataset is utilized which includes approximately 14,700 images divided into 49 classes. The proposed QMX model achieves an overall F1 accuracy of 98%. In addition, a comprehensive analysis among QMX, Modified Xception Tiny (MXT), MX, and the default Xception model is provided in this research. Finally, the model has been implemented on Raspberry Pi 4 and a detailed evaluation of its performance has been conducted, including a comparison with existing state-of-the-art approaches in this domain. The results demonstrate that the proposed QMX model outperforms the prior work in terms of performance. Keywords—Bengali sign language; CNN; computer vision; model quantization; Raspberry Pi 4; transfer learning; Tiny ML I. I NTRODUCTION Disability is a crucial issue in terms of human rights because a person with an impairment is usually deprived of ordinary public welfare. Almost a billion of the world’s popu- lation has some form of physical disability 1 . Individuals with disabilities experience more negative socioeconomic conse- quences, resulting in a poorer standard of life. While over 430 million people worldwide suffer from hearing impairment 2 , there are more than 1.7 million hearing and speaking impaired people in Bangladesh alone 3 . These impaired people belong to 1 https://www.who.int/news-room/fact-sheets/detail/disability-and-health 2 https://www.who.int/news-room/fact-sheets/detail/deafness-and-hearing- loss 3 https://en.wikipedia.org/wiki/Deafness in Bangladesh the Bangladesh Deaf and Mute Community (BDMC). Due to their communication impediment, the BDMC, faces numerous obstacles while attempting to participate in education, work, social activities, and other aspects of everyday life. Sign language employs the visual-manual paradigm to communicate meaning. Sign language is conveyed via hand and finger movements to create gestures. The only way to communicate with people with hearing or speaking disabilities is through sign language. Similar to every other language, the Bengali language has its own sign language, which is known as Bengali Sign Language (BdSL). The BDMC uses only BdSL to communicate with everyone, which restricts their ability to converse with society, as the majority of the society does not know sign language due to a lack of social awareness. In the aforementioned scenario, communication between the BDMC and society requires a sign language interpreter. However, a skilled interpreter may not always be readily available, and in such circumstances, paying fair fees may be a serious worry. An automated recognition system for sign language can play a vital role in reducing the basic and social differences between society and BDMC. Therefore, sign language recognition is a popular area of study. Current research in this area focuses mostly on either sensor-based [1] or vision-based [2] systems. Numerous studies have been conducted on BdSL recogni- tion, and there are numerous benchmarking datasets [3], [4], [5], [6], [7] for BdSL recognition. However, these datasets are insufficient for training and evaluating deep learning models, and the majority are not open-source. CNN [8], [9] is a popular choice along with the transfer learning [10], [11], [12] model to recognize BdSL. Several research implements the CNN model for recog- nizing BdSL. Hossain et al. [13] proposed a CNN-based sign language recognition model and achieved 98.75% accuracy. Islalm et al. [14] also proposed a CNN-based model, and they evaluated their model using 10-fold cross-validation. They achieved 99.80% accuracy. Some research utilized CNN-based transfer learning models for recognition. Rafi et al. [4] utilized the VGG19 transfer learning model with 89.6% accuracy. To our knowledge, no prior work exists on constructing an efficient deep learning model that can be implemented in embedded or IoT devices via model quantization. The majority of recent work employed mainstream or pre-trained models. Therefore, these models cannot be implemented on devices with a low configuration. www.ijacsa.thesai.org 1099 | Page