Just Recognizable Distortion for Machine Vision Oriented Image and Video Coding

摘要

Machine visual intelligence has exploded in recent years. Large-scale, high-quality image and video datasets significantly empower learning-based machine vision models, especially deep-learning models. However, images and videos are usually compressed before being analyzed in practical situations where transmission or storage is limited, leading to a noticeable performance loss of vision models. In this work, we broadly investigate the impact on the performance of machine vision from image and video coding. Based on the investigation, we propose Just Recognizable Distortion (JRD) to present the maximum distortion caused by data compression that will reduce the machine vision model performance to an unacceptable level. A large-scale JRD-annotated dataset containing over 340,000 images is built for various machine vision tasks, where the factors for different JRDs are studied. Furthermore, an ensemble-learning-based framework is established to predict the JRDs for diverse vision tasks under few- and non-reference conditions, which consists of multiple binary classifiers to improve the prediction accuracy. Experiments prove the effectiveness of the proposed JRD-guided image and video coding to significantly improve compression and machine vision performance. Applying predicted JRD is able to achieve remarkably better machine vision task accuracy and save a large number of bits.