DeReFNet: Dual-stream Dense Residual Fusion Network for static hand gesture recognition

• Published in: Displays Vol. 77; p. 102388
• Main Authors: Sahoo, Jaya Prakash, Sahoo, Suraj Prakash, Ari, Samit, Patra, Sarat Kumar
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.04.2023
• Subjects: Convolutional neural network; Dense connectivity; Feature concatenation module; Global feature aggregation block; Hand gesture recognition; Residual network; Global feature aggregation block; Feature concatenation module; Convolutional neural network; Residual network; Dense connectivity; Hand gesture recognition
• ISSN: 0141-9382; 1872-7387
• DOI: 10.1016/j.displa.2023.102388

Vision-based hand gesture recognition (HGR) system provides the most effective and natural way of interaction between humans and machines. However, the recognition performance of such an HGR system is challenging due to the variations in illumination, complex backgrounds, the shape of the user’s hand, and inter-class similarity. This work proposes a compact dual-stream dense residual fusion network (DeReFNet) to address the above challenges. The proposed convolutional neural network architecture mainly utilizes the strength of global features from each residual block of the residual stream and spatial information from the other stream using dense connectivity. Both the streams are fused to gather enriched information using the feature concatenation module. The efficacy of the DeReFNet is validated using a subject-independent cross-validation technique on four publicly available benchmark datasets. Furthermore, the qualitative and quantitative analysis of the benchmarked datasets illustrates that the DeReFNet outperforms state-of-the-art methods in terms of accuracy and computational time. •Development of a dual-stream deep network for accurate recognition of hand gestures.•The first residual stream extracts global information from each residual block.•The second stream extracts spatial features through dense connectivity.•The proposed network has less trainable parameters leading to lower inference time.•Our network outperforms the existing techniques in computational time and accuracy.