ADJSCC-l: SNR-Adaptive JSCC Networks for Multi-Layer Wireless Image Transmission
Xiuwen Bao, Ming Jiang, Hao Zhang
Abstract
Motivated by deep learning in wireless image transmissions, joint source-channel coding (JSCC) techniques based on convolutional neural networks (CNNs) can successfully improve the image quality in noisy channels compared with conventional digital techniques. In urgent or limited scenarios like the surveillance communication, image information with a low compression rate is sent at the first transmission stage to save the channel rescources. If the quality of the reconstruction is of low-quality, the supplementary information can be transmitted at the second stage to improve the quality of the reconstruction. In this paper, we propose an attention mechanism based multi-layer JSCC architecture for the progressive image transmission, called ADJSCC-l. In the proposed architecture, the first layer can recover the base information with the highest compression rate in the first transmission, while the refinement layers can utilize the refinement information in successive transmissions to obtain better image quality.The provided numerical results show that the ADJSCC-l offers a bi-mode transmission which can significantly enhance the validity and reliability of the data in wireless image transmission compared with other neural network (NN) based JSCC schemes.