Optimization of Compression Rate Allocation for Image Segmentation Regions in MIMO DeepJSCC
Goki Sawada, Shion Inokuma, Daisuke Hisano, Kazuki Maruta · 2025
This paper proposes optimizing the compression rate of segmented image region in deep joint source-channel coding (DeepJSCC) transmission in order to improve both image quality and communication efficiency. When transmitting highresolution images with DeepJSCC, segmented transmission is employed due to computational constraints. Additionally, the structure of the output layer of the convolutional neural network (CNN) can adjust its number of transmitted symbols for each region. Focusing on this feature, we attempt to optimize the compression rate for each region of the image based on the segment's entropy value. Its validity is examined through singleinput single-output (SISO) and multiple-input multiple-output (MIMO) channels demonstrating improvements in decoding accuracy of target regions while maintaining overall image quality, especially in low-SNR environments. Beyond the entropy-based importance classification used in this paper, the potential for further accuracy improvements through more precise importance classification is also suggested.