Fig. 1. Technical roadmap of end-to-end learning-based image compression

Fig. 2. Nonlinear transform based end-to-end learning image compression^[10]

Fig. 3. Images obtained by different quantization methods using JPEG compression^[14]. (a) Original image; (b) rounding; (c) stochastic rounding; (d) additive noise

Fig. 4. Autoencoder based on hyperprior and recursive nearest neighbor probability fusion^[27]

Fig. 5. Objective evaluation. (a) Pixel-level distortion measured by MS-SSIM (dB); (b) PSNR used for structural similarity evaluation

Fig. 6. Subjective evaluation. (a) JPEG420; (b) BPG444; (c) NLAIC MSE opt; (d) NLAIC MS-SSIM opt;(e) original image