DBL-SC: background-independent sign language recognition based on spatial channel separation computation
摘要
Sign language recognition aims to bridge the communication gap between the hearing-impaired and hearing individuals. However, existing methods have been proposed on pure background datasets, making it challenging to apply them in real-world scenarios. One of the primary reasons for this challenge is the complex and variable backgrounds that affect the distribution of data in both the training and test sets. Therefore, it is essential to investigate how to apply models trained on pure backgrounds to complex real backgrounds. To address this, we propose a spatial channel separation computation (DBL-SC) sign language recognition method to mitigate the influence of background features on recognition. Specifically, our proposed DBL-SC uses group attention and channel dropout to address the different distribution forms of background noise in low-level and high-level features. Additionally, we build a background replacement continuous sign language (BR-CSL) dataset for the experiment. The training set contains sign language videos with pure backgrounds, while the test set includes sign language videos with real backgrounds. Our experiments on this dataset show a significant improvement of 7.96