Learned Image Coding for Human-Machine Collaborative Optimization

The exponential growth in the volume of image data has imposed immense pressure on transmission and storage systems, while simultaneously presenting opportunities for intelligent image analysis towards machine vision. Recent years, learned image coding approach have made remarkable advancements with impressive performance. The application of the learned image coding method in machine vision holds promising prospects for achieving human-machine collaboration. In this paper, we propose a learned image coding approach based on Transformer-CNN interaction structure for human-machine vision collaborative optimization, which can generate a single and compact bitstream for efficient representation in image compression. The bitstream can be directly decoded to generate a reconstructed image for human visual perception. In parallel, without the need for decoding and reconstructing the image, the bitstream can serve as input for machine vision tasks. This not only reduces computational costs on the decoding end but also enhances machine analysis efficiency. Experimental results demonstrate that our proposed learned image coding method achieves a single bitstream that concurrently considers image reconstruction and machine task analysis, ensuring high accuracy in machine tasks and superior quality in reconstructed images compared to state-of-the-art (SOTA) methods.