Scholarlink

<p>People with difficulty hearing or speaking are heavily dependent on nonverbal, hand-gestation-based forms of communication. Sign language is a significant source of communication for deaf and hard-of-hearing people. Therefore, implementing a system that can recognize sign language would significantly assist dumb, deaf, and hard-of-hearing individuals. Sign language recognition (SLR) has been a highly active area of research in recent years. Finding an accurate result from different image datasets is critical for an automatic SLR system. Although convolutional neural networks (CNNs) have shown outstanding performance, there is still a possibility for improvement, particularly when dealing with multiple resources and algorithms. This paper proposes a convolutional neural mixed approach model (CNMAMs) based on a traditional CNN model with different filters and extra-layered architecture. The Sobel gradient-based filter is utilized to detect edges by highlighting areas with high spatial frequency, and we also employ a Gabor filter to identify specific frequencies and orientations for making it effective for texture analysis and angled edge detection. Additional convolutional layers can help the model learn more complex features. We add layers with smaller filters(3X3) after existing layers to capture finer details. Similarly, adding an additional batch normalization layer after convolutional layers can help stabilize the learning process and allow for higher learning rates to our model. This model can process various data sets and is useful for streaming real-time images. The performance of the proposed methods is evaluated by conducting several experiments on real-time images and different benchmark datasets: the Indian sign language (ISL) and American sign language (ASL) datasets. The accuracy achieved for the double-handed gesture was 97.91%, and for the single-handed gesture, it was 96.70%. The experimental results indicate that the proposed model outperforms the existing approaches in sign language recognition. The proposed model can be further extended to increase the use of SLR in diverse sign language datasets. The movement of the hand and face can also be taken into consideration for recognition.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid Convolutional Neural Mixed Approached Model for Incorporating Sign Language Features

  • 期刊论文

  • 发表日期：

    2025年6月23日

查看全文
  • Varun Prakash Saxena,
  • Premal P. Shah,
  • Malaram Kumhar,
  • Jitendra Bhatia,
  • Prafulla Saxena,
  • Sudeep Tanwar

摘要

People with difficulty hearing or speaking are heavily dependent on nonverbal, hand-gestation-based forms of communication. Sign language is a significant source of communication for deaf and hard-of-hearing people. Therefore, implementing a system that can recognize sign language would significantly assist dumb, deaf, and hard-of-hearing individuals. Sign language recognition (SLR) has been a highly active area of research in recent years. Finding an accurate result from different image datasets is critical for an automatic SLR system. Although convolutional neural networks (CNNs) have shown outstanding performance, there is still a possibility for improvement, particularly when dealing with multiple resources and algorithms. This paper proposes a convolutional neural mixed approach model (CNMAMs) based on a traditional CNN model with different filters and extra-layered architecture. The Sobel gradient-based filter is utilized to detect edges by highlighting areas with high spatial frequency, and we also employ a Gabor filter to identify specific frequencies and orientations for making it effective for texture analysis and angled edge detection. Additional convolutional layers can help the model learn more complex features. We add layers with smaller filters(3X3) after existing layers to capture finer details. Similarly, adding an additional batch normalization layer after convolutional layers can help stabilize the learning process and allow for higher learning rates to our model. This model can process various data sets and is useful for streaming real-time images. The performance of the proposed methods is evaluated by conducting several experiments on real-time images and different benchmark datasets: the Indian sign language (ISL) and American sign language (ASL) datasets. The accuracy achieved for the double-handed gesture was 97.91%, and for the single-handed gesture, it was 96.70%. The experimental results indicate that the proposed model outperforms the existing approaches in sign language recognition. The proposed model can be further extended to increase the use of SLR in diverse sign language datasets. The movement of the hand and face can also be taken into consideration for recognition.