An Implementation of Picture Compression with A CNN-based Auto-encoder

Ming Li, Jianhua Hu, Changsheng Xia, Yundong Zhang; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018, pp. 2543-2546

We mainly use the importance-map CNN method introduced by Mu.Li[??] to compress the CLIC2018 validation and test pictures. The framework is an autoencoder, with the bottleneck containing a 4-bit importance map and a 1/8 scale-down feature maps(FMs) of 64-channel and 1-bit contents. We re-implemented this model in the Tensorflow/python enviroment. Different from the original work, we modify the network a little to ge better performance and creatively replace the entropy-coding scheme with a much simpler reorder and run-length coding method. We also share some techniques and experiences for model training and fine tuning the encoder for the CLIC2018 test pictures. Method of controlling the final bit rate is also mentioned.