Deep Joint Source-Channel Coding Using FFT-Enabled Convolutional Neural Network for Image Transmission

Deep joint source-channel coding (DeepJSCC) has attracted attention as a type of semantic communication that shares not only information but also meaning and intent, and it is a type of deep learning that uses an autoencoder instead of conventional source and channel coding. DeepJSCC with convolutional neural network (CNN) layers has been studied specifically for image transmission, where it avoids cliff effects and achieves higher image quality even in the low signal-to-noise power ratio (SNR) range. In addition, DeepJSCC is suit for services related to the Internet of things (IoT), and is likely to be widely applicable to monitoring-related applications such as surveillance cameras. For this reason, it may not be possible to provide sufficient GPU resources, especially for the transmitter where the encoder is installed. Therefore, we propose a fast Fourier transform (FFT)-based DeepJSCC that replaces CNN with FFT and element-wise product to reduce the computation time of DeepJSCC. This paper evaluates the impact of changing the layer structure on image quality.