A Joint Optimization Framework for Codec Models Targeting Machine Vision Tasks

With the rapid advancement and improved reliability of machine vision technologies, tasks such as license plate recognition and industrial defect detection can now rely almost entirely on automated visual systems, significantly reducing labor costs. However, how to better adapt video compression algorithms for machine vision tasks remains an urgent challenge. To address this, we propose a joint training method tailored for machine vision. The method incorporates deep feature distortion loss from a downstream task network into the optimization of the codec, enabling the model to preserve information more relevant to machine vision during compression. Additionally, we introduce a magnitude-balanced adaptive weighted averaging algorithm to ensure that the distortion loss of the reconstructed image and that of the feature space are learned at similar rates. Experimental results demonstrate that, compared with conventional training methods, the proposed approach achieves a 22.22% improvement in BD-rate, making it more suitable for training codecs oriented towards machine vision tasks.