Research On Joint Source Channel Coding For Task Classfication

In recent years,due to the continuous development of hardware,processor performance has been greatly improved,which greatly improves the development speed of artificial intelligence,at the same time,communication technology is also constantly upgraded,it can be predicted that the future integration of artificial intelligence and communication development will be more and more close.Therefore,it is of certain potential application value to study the intersection of artificial intelligence and communication technology.In this paper,a deep joint source channel coding scheme is designed for the classification task in computer vision.Aiming at the classification task that the receiver does not need to reconstruct the source,this paper improves the existing deep joint source channel coding technique and introduces a strategy mechanism in the communication sender.Driven by classification tasks,the policy mechanism can learn the features of source data,and the degree of feature compression can be changed by adjusting the ratio factor of the policy mechanism in the loss function.Compared with the existing deep joint source channel coding schemes,the classification performance of the deep joint source channel coding scheme introduced in this paper is greatly improved when the same feature quantity is transmitted in AWGN channel.In this paper,a deep joint source channel coding scheme based on anti-noise autoencoder is designed for the task that the receiver needs to reconstruct the source classification.In order to improve the reconstruction performance and classification performance of the receiver,the transmitter improves the deep joint source channel coding scheme with attention mechanism,and integrates the channel noise into the features extracted by the coding network to improve the anti-noise capability of the deep joint source channel coding network.Both the proposed scheme and the baseline scheme trained under fixed signal-to-noise ratio(SNR)without introducing attention mechanism are better than the baseline scheme when tested in AWGN channel.Compared with the improved attention mechanism,the proposed scheme has improved the classification performance while maintaining the same reconstruction performance.