信号处理2025，Vol.41Issue(2)：399-408,10.DOI:10.12466/xhcl.2025.02.017
任务自适应增强的人机特征解耦可分级压缩
Scalable Image Coding via Feature Decoupling for Human and Machine with Task-Adaptive Enhancement
摘要
Abstract
Image compression is a critical technology designed to minimize information redundancy during transmission while preserving the quality of the compressed image.With advancements in computer vision,images are increasingly consumed by machines in addition to humans,necessitating compression methods that cater to both human and machine vision requirements.While current learning-based image coding techniques have significantly improved human visual perception,they struggle to balance signal fidelity and semantic fidelity,limiting their ability to meet the needs of both audiences effectively.To address this limitation,this study proposes a task-adaptive,feature-decoupling scalable com-pression method.This approach supports multiple machine vision tasks using a single bitstream and enables selective or complete image reconstruction depending on specific requirements.The proposed method decouples image features into object and background features,compressing and reconstructing them independently.Reconstructed object features are employed for tasks such as object detection and semantic segmentation,whereas the fully reconstructed image caters to human visual perception.This method enhances compression efficiency by eliminating the need to reconstruct the entire image for visual tasks,thereby meeting the distinct demands of human perception.Furthermore,to address performance imbalances arising from variations in the importance of target regions,a plug-and-play task-adaptive unit is integrated into the target feature decoder.This unit enables task-specific adjustments to improve the analysis performance of recon-structed target images without requiring retraining of the entire network.Experimental results demonstrate that the pro-posed method outperforms conventional encoders and decoders in task performance while achieving superior Rate-Distortion efficiency.These findings underscore the potential of this method to advance scalable image compression for both human and machine vision applications.关键词
图像压缩/人机协同/特征解耦/任务自适应增强Key words
image compression/human-machine collaborative/feature decoupling/task-adaptive enhancement分类
信息技术与安全科学引用本文复制引用
安平,沙莉娅,吴颖,杨超,黄新彭..任务自适应增强的人机特征解耦可分级压缩[J].信号处理,2025,41(2):399-408,10.基金项目
国家自然科学基金(62071287,62020106011,62371279,62371278) （62071287,62020106011,62371279,62371278）
上海市科学技术委员会基金(22ZR1424300)The National Natural Science Foundation of China(62071287,62020106011,62371279,62371278) （22ZR1424300）
Science and Technology Commission of Shanghai Municipality(22ZR1424300) （22ZR1424300）
