Abstract
Machine-oriented image compression has become a critical challenge in large-scale surveillance networks, where camera-captured images are compressed through a codec before being forwarded to server-side analytics pipelines. However, codecs optimized for human perception routinely discard the semantic cues that machine vision models rely on, degrading downstream task performance. To bridge this gap, existing methods take two directions. Some methods embed trainable modules into a specific LIC architecture, limiting deployment to LIC pipelines, while others prepend a neural pre-processor to the codec but overfit to the proxy codec during training. Codec transferability thus remains an unaddressed limitation, critical for surveillance systems where multiple conventional codecs are deployed across heterogeneous camera networks. In this paper, we propose Universal Image Codec Adapter for Machine vision (UICAM), a lightweight codec-transferable adapter that addresses the proxy-codec overfitting inherent to neural pre-processor approaches. We introduce a post-codec decoder that reshapes the backward gradient during training, expanding the optimization directions available to the pre-codec encoder beyond the proxy codec’s distortion patterns and enabling transfer to any unseen codec without fine-tuning. Additionally, UICAM is trained with a dual-granularity perceptual loss (DGPL) that jointly supervises global and dense semantics, enabling a single model to cover multiple downstream tasks. UICAM achieves up to 85% BD-Rate reduction on unseen LICs and 75% on conventional codecs across six downstream tasks, using 9× to 19× fewer parameters than prior pre-processor methods.
Seongmoon Jeong (정성문)
Combined MS-PhD student
