Chinese sign language recognition method based on a variational encoder

Abstract

The invention relates to a Chinese sign language recognition method based on a variational encoder. The Chinese sign language recognition method comprises the steps that firstly collecting and reconstructing time series data of the Chinese sign language by the encoder, and measuring the performance of the encoder with relative entropy, which means measuring KL divergence of the encoder; constructing the encoder with LSTM type RNN according to the characteristic of the time series to reconstruct eigenvector of the sign language time series; then using a unsupervised learning mode to reversely decode the reconstructed data; constructing a decoder with LSTM type RNN; finally calculating cross entropy between input data and decoded output data to get loss function of the entire structure, passing back the error, constantly updating the encoder and the decoder parameters, minimizing the loss function, so as to obtain a final CODEC model for sign language recognition. The Chinese sign language recognition method can reduce the interference of nonlinear disturbance signal, realize the unsupervised learning, simplify the time series identification network and improve the accuracy rate of the Chinese continuous sign language recognition.

Description

technical field [0001] The invention relates to the field of Chinese sign language recognition, in particular to a Chinese sign language recognition method based on a variable coder. Background technique [0002] Sign language recognition is a technology that can convert sign language information into voice, text, and read or display. In the field of sign language recognition, since continuous sign language recognition is the key issue of sign language recognition, the key to how to improve the effect of sign language recognition is how to improve the accuracy of continuous sign language recognition. [0003] In the prior art, there are mainly the following methods for continuous sign language recognition: [0004] First, continuous sign language recognition usually uses HMM (Hidden Markov Model, Hidden Markov). This method introduces the influence of the previous state on the current state in the model, and realizes sign language recognition by maximizing the calculated ou...

AI Technical Summary

Problems solved by technology

[0008] Although the method of left and right expansion can introduce the relationship between the front and rear states to a certain extent, but in order to reduce the size and complexity of the model, the expansion size is very limited, so the distance before and after the link should not be too far, resulting in the decline of the ability to perceive the previous state at the current moment ;

[0009] Supervised learning requires manual data labeling, cumbersome data collection and heavy workload

[0010] The impact of nonlinear disturbances on the recognition results is not considered. When the data has small disturbances, the robustness of the recognition algorithm is not strong

Images