Adaptable Semantic Compression and Transmission Resource Allocation for Intelligent Image Classification

With the development of artificial intelligence, task-oriented semantic communication with low latency and high accuracy is a new communication mode for edge-aided Internet of Things (IoT). In this paper, we propose an image classification-oriented semantic communication method for multi-user IoT system, which aims to implement the adaptable semantic compression (ASC) and optimize the transmission resource allocations based on the semantic relationships. Specifically, to satisfy the low-latency and high-accuracy requirement, the semantic entropy of the transmitted image multiplied by the success transmission probability is considered as the success probability of classification tasks, which is to be maximized given fixed total transmission power and bandwidth. In order to solve this maximization issue, we use the particle swarm optimization algorithm to optimize the compression ratio and the allocation of power and bandwidth for each user jointly. Simulation results show that the proposed ASC-based semantic scheme has much better performance in image transmission and classification, and is more robust under different channel conditions. Without affecting the probability of successful task completion, the proposed method can reduce the size of the transmitted data by 40% ~ 80%. It is verified that a reasonable selection of compression ratio and intelligent allocations of communication resources for different users not only guarantee the high accuracy of the image classification, but also further reduce the transmission delay, which is of great practical interest.