Densely connected convolutional network block based autoencoder for panorama map compression

As a novel virtual reality (VR) format, panorama maps are attracting increasing attention, while the compression of panorama images is still a concern. In this paper, a densely connected convolutional network block (dense block) based autoencoder is proposed to compress panorama maps. In the proposed autoencoder, dense blocks are specially designed to reuse feature maps and reduce redundancy of features. Meanwhile, a loss function, which imports a position-dependent weight item for each pixel, is proposed to train and adjust network parameters, in order to make the autoencoder fit to properties of panorama maps. Based on the proposed autoencoder and the weighted loss function, a greedy block-wise training scheme is also designed to avoid gradient vanishing problem and speed up training. During training process, the autoencoder is divided into several sub-nets. After each sub-net is trained separately, the whole network is fine-tuned to achieve the best performance. Experimental results demonstrate that the proposed autoencoder, compared with JPEG, saves up to 79.69 % bit rates, and obtains 7.27dB gain in BD-WS-PSNR or 0.0789 gain in BD-WS-SSIM. The proposed autoencoder also outperforms JPEG 2000, HEVC and VVC in both BD-WS-PSNR and BD-WS-SSIM. Meanwhile, subjective results show that the proposed autoencoder can recover details of panorama images, and reconstruct maps with high visual quality.