Activation Map Saliency Guided Filtering for Efficient Image Compression for Vision Tasks

With the development of Internet of things applications, there will be ever more visual data generated for machine vision tasks. Traditional image compression, which focus on recovering pixels from compressed representations for human consumption, could waste bits on non-essential information, as the compression pipeline is designed around a pixel-level loss function. How to design new compression schemes that directly reflect vision task losses, becomes the key in extracting new compression efficiency in machine vision applications. Recent research in the field of machine learning understanding provides us activation map saliency as a kind of deep in-sight to deep-learning-based inference. In this paper, we propose an image compression scheme based on activation map guided filtering to encode images with higher compression ratio while maintaining classification accuracy. Specifically, a pre-filter is adopted at encoder side to improve traditional image compression standards. This prefilter is designed to preserve value and edge information of salient area for analysis and smooth the unimportant pixels for bits saving. The images compressed by our scheme could be decoded for vision task without any modification to standard decoder. The experimental results show that, compared with traditional compression methods, our method could improve compression efficiency while maintaining the classification accuracy.