D-CBAMFi: A Lightweight Multi-Scale Fusion Network Based on Depthwise Separable Convolution and Dilated Convolution

Gesture recognition is crucial for the development of new human-computer interaction paradigms that will change the interaction between humans and their surroundings. Most deep learning-based techniques use neural networks with a single backbone architecture, which can only capture a single feature and cannot meet the requirements of real-world applications for recognition tasks in resource-limited and complex scenarios. Therefore, we propose a lightweight multi-scale fusion network, D-CBAMFi, whose core component is a D-CBAM network consisting of parallel depth-wise separable convolution and dilated convolution and CBAM. The proposed system achieves an average accuracy of 84.30% on the Widar dataset, validating the effectiveness of the model.