Adaptive Image Downscaling for Rate-Accuracy-Latency Optimization of Task-Target Image Compression

Hangyul Choi, Seongmoon Jeong, Sangwoon Kwak, Soon-Heung Jung, Jong Hwan Ko

Abstract: As the computational cost of the compression frame-work increases due to improved image codecs and the proliferation of high-resolution images, downscaling can be used to reduce compression overhead before compression process. There are several existing deep learning-based downscaling approaches, but these studies have the limitation of outputting fixed sizes for all different input images. In this paper, we propose a novel approach of adaptive image downscaling framework for rate-accuracy-latency optimization. We utilize deep learning-based downscaling network that learns which size factor to use in downscaling operation adjusting the trade-off between rate and accuracy through λ. Our experimental results show that the proposed framework enhances the rate-accuracy performance of compression rate control or uniform downscaling by up to 43.5% BD-rate (mAP), while remaining minimal latency at the same accuracy compared to others.