Feature Importance-Aware Deep Joint Source-Channel Coding for Computationally Efficient and Adjustable Image Transmission

Recent advancements in deep learning-based joint source-channel coding (deepJSCC) have significantly improved communication performance, but their high computational demands restrict practical deployment. Furthermore, some applications require the adaptive adjustment of computational complexity. To address these challenges, we propose a computationally efficient and adjustable deepJSCC model for image transmission, which we call feature importance-aware deepJSCC (FAJSCC). Unlike existing deepJSCC models that equally process all neural features of images, FAJSCC first classifies features into important and less important features and then processes them differently. Specifically, computationally-intensive self-attention is applied to the important features and computationally-efficient spatial attention to the less important ones. The feature classification is based on the available computational budget and importance scores predicted by an importance predictor, which estimates each feature's contribution to performance. It also allows independent adjustment of encoder and decoder complexity within a single trained model. With these properties, our FAJSCC is the first deepJSCC that is computationally efficient and adjustable while maintaining high performance. Experiments demonstrate that our FAJSCC achieves higher image transmission performance across various channel conditions while using less computational complexity than the recent state-of-the-art models. Adding to this, by separately varying the computational resources of the encoder and decoder, it is concluded that the decoder's error correction function requires the largest computational complexity in FAJSCC, which is the first observation in deepJSCC literature. The FAJSCC code is publicly available at https://github.com/hansung-choi/FAJSCC.