This paper describes the technology proposal by NCTU for learning-based image compression. The selected technologies include an autoencoder that incorporates (1) a principal component analysis (PCA) layer for energy compaction, (2) a uniform, scalar quantizer for lossy compression, (3) a context-adaptive bitplane coder for entropy coding, and (4) a soft-bit-based rate estimator. The PCA layer includes 1×1 eigen kernels derived from the sample covariance of co-located feature samples across channels. The bitplane coder compresses PCA-transformed feature samples based on their quantized, fixed-point representations, of which the soft bits provide a differentiable approximation for context-adaptive rate estimation. The training of our compression system proceeds in two alternating phases: one for updating the rate estimator and the other for fine tuning the autoencoder regularized by the rate estimator. The proposed method outperforms BPG in terms of both PSNR and MS-SSIM. Several bug fixes have been made since the submission of our decoder. This paper presents the up-to-date results.