Task-Driven Semantic-Aware Image Compression

In the context of IoT/IoE, efficient image compression is crucial for reducing bandwidth consumption and enabling real-time data processing. Predominantly, existing compression techniques are designed with an emphasis on human visual perception, frequently overlooking the specific demands of tasks involving image analysis. This paper introduces a novel Task-Driven Semantic-Aware Image Compression (TDSIC) approach with a feature enhancement network engineered to retrieve the latent semantic information typically lost in entropy-decoded results. TDSIC also leverages importance maps as a form of prior knowledge, which assists in guiding the bit allocation strategy of the compression technique. By refining the perceptual loss function, the network’s emphasis is realigned to prioritize the retention of semantically significant information. Additionally, the training regimen incorporates insights from discriminative networks to harness feedback on human visual discernment, which is instrumental in meticulously calibrating the perceptual understanding of the image compression model. This calibration ensures that the model adeptly captures and correlates pertinent information. TDSIC achieves up to a 12% improvement in Top-1 image classification accuracy on the ILSVRC2012 dataset and up to a 0.3% enhancement in object detection on the COCO dataset over baseline methods. The Kodak dataset experiments underscore that images compressed with our method not only maintain visual appeal to human viewers but also enhance the performance of specific tasks. Supplementary investigations, including ablation studies on the feature enhancement module (FEM), indicate that the FEM is adept at recovering information that was previously lost during the encoding and quantization phases. The implementation code is made publicly accessible at https://github.com/lie209/TDSIC.