关键词: 手势识别, 注意力机制, 复杂背景, 结构重参数化, 数据增强

Abstract:

As a highly adaptive form of interaction in human-computer interaction, gestures can simplify interactions by eliminating physical contacts between mechanical devices and their users. Gesture interaction provides more intuitive interaction and richer interaction effects, better meeting people’s needs and expectations for interaction. Gesture recognition has been widely researched in the field of human-computer interaction, especially gesture recognition based on machine vision thanks to its low cost, being more natural and non-contact. However, the existing gesture recognition methods are primarily based on simple experimental environment background. In the actual human-computer interaction, gesture recognition usually occurs in various complex environments.

In practice, changes in brightness, complex backgrounds, and color-like interference are key factors affecting the accuracy of gesture recognition. The interference caused by complex background greatly affects the extraction of gesture features, making it difficult to recognize gestures quickly and accurately. Some researchers employ a two-stage model to first extract gesture areas and then identify them, while others directly use deep convolutional neural networks to identify complex background gestures. However, the recognition speed of the two-stage gesture recognition method hardly meets the requirements in practical applications, and the accuracy of the single-stage gesture recognition method needs to be further improved for the gesture image recognition of complex background. The existing gesture recognition methods are unable to solve the problems of gesture recognition in the actual complex background due to their difficulties in striking a balance between recognition speed and accuracy. To remedy this, the key lies in how to eliminate or weaken the interference of complex background on the basis of improving the recognition speed of the algorithm, or how to enhance the ability of gesture feature extraction, so that the gesture recognition algorithm can correctly represent the gesture information. The attention mechanism can imitate the principle of human visual system’s attention to objects, by increasing the attention to the target area to achieve the detailed information of the target area. Embedding attention mechanism in gesture recognition algorithm based on deep learning can allow the algorithm to focus on the feature of target gesture area and eliminate the interference of complex background. Meanwhile, the structure reparameterization method can remove the redundant branch structure in the deployment stage and improve the algorithm recognition speed.

To remedy such problems as low recognition accuracy and slow recognition speed caused by more interference in gesture images under complex background, a gesture recognition algorithm RepSEHGR based on structural reparameterization and attention mechanism is proposed. By using the structure reparameterization method, it is applied to the residual structure to remove the redundant branch structure in the deployment stage and improve the algorithm recognition speed. Meanwhile, the channel attention mechanism module is embedded to enable the algorithm to attend to gesture features by weighted features of different channels, thus reducing complex background interference. Finally, two data enhancement methods, cutout and affine transformation, are employed to train the algorithm, suppress complex background noise input and enhance the data, reduce overfitting and improve the robustness of the algorithm. Comparison experiments on a complex background gesture data set show the recognition accuracy reaches 99.9% and the recognition speed 200FPS, demonstrating the effectiveness of the proposed algorithm.