Deep learning-based joint source-channel coding (Deep JSCC) has attracted attention. Deep JSCC maps information source features directly to IQ symbols using an autoencoder instead of source and channel coding. In image transmission, severe image degradation due to the cliff effect has been a problem in the lower signal-to-noise ratio (SNR) regions. Meanwhile, Deep JSCC provides a better peak signal-to-noise ratio (PSNR) even in low SNR regions. However, a high-resolution image needs to be divided into small patch images to input into Deep JSCC, consideration of the limitation of transmission capacity and delay. The image division causes block noise and PSNR degradation. This paper proposes Deep JSCC with overlapping image division to suppress the block noise. The proposed overlap division overlaps several pixels between patch images when dividing into patch images. After image recomposition by the decoder of Deep JSCC, the receiver removes the overlapping pixel information and combines the edge of the patch images smoothly. This paper conducts the experiments to reveal the effectiveness of block noise reduction owing to the proposed scheme with DIV2K data set. Consequently, we indicate the PSNR is drastically improved in the specific SNR region.
Deep Joint Source-Channel Coding Using Overlap Image Division for Block Noise Reduction
2024-06-24
2469217 byte
Conference paper
Electronic Resource
English
Image transmission over multipath fading channel using joint source channel coding modulation
British Library Online Contents | 2009
|