A Lossy Image Compression Algorithm Based On Fusion Of Multi-scale Cascaded Neural Networks And Discrete Cosine Transform

This paper introduces an efficient lossy image compression algorithm which combines multi-scale cascaded neural networks with Discrete Cosine Transform (DCT), in response to MPEG VCM Call for Evidence. The multi-scale cascaded neural networks are firstly used to compress the features and obtain a low-dimensional compact representation. Although this low-dimensional compact representation loses feature details, it capturing several key important cues within the feature at an extremely low bit-rate cost. Then, DCT is used to encode and decode the condensed features. At last, the decoded features produced by DCT are restored by the multi-scale cascaded neural networks. The proposed image compression algorithm is evaluated on multi-object tracking task and the experimental results show that this method just needs much fewer time to compress image at the cost of slightly sacrificing image quality in comparison to the standard methods given by MEPG.