Content Adaptive Multi-Scale Feature Layer Filtering

Abstract:

This paper presents a content-adaptive feature layer filtering method for intermediate feature compression in split inference systems using multi-scale neural networks. The proposed encoder-side optimization removes redundant feature layers based on object size information derived from the input image. Early layers, which contain high spatial resolution within feature maps are suited for detecting small objects. These early layers are then pruned when large objects dominate the scene and their contribution becomes negligible. This reduces redundancy and improves compression efficiency. The method requires no retraining of the task network and remains compatible with conventional codecs by spatially packing the retained features. Aligned with the MPEG Feature Coding for Machines (FCM) framework, this approach enables more efficient collaborative intelligence by reducing bandwidth during intermediate feature transmission. Experimental results on object detection and segmentation tasks show up to a 43% bitrate reduction without compromising task accuracy.

Authors:

Juan Merlos, Md Eimran Hossain Eimon, Ashan Perera, Hari Kalva, Velibor Adzic, Borko Furht

Conference / Journal

2025 IEEE 8th International Conference on Multimedia Information Processing and Retrieval (MIPR)