Residual‐error prediction based on deep learning for lossless image compression

Abstract

A novel residual-error prediction method based on deep learning with application in lossless image compression is introduced. The proposed method employs machine learning tools to minimise the residual error of the employed prediction tools. Experimental results demonstrate average bitrate savings of 32% over the state-of-the-art in lossless image compression. To the best of the authors' knowledge, this Letter is the first to propose a deep-learning based method for residual-error prediction.

Deep-learning-based macro-pixel synthesis and lossless coding of light field images

“…(i) 29.9 improvement compared to CALIC [2], a traditional lossless image codec; (ii) 12.5 improvement compared to our previous method [6]; (iii) 9.1 improvement compared to FLIF [3], the current state-of-the-art lossless image codec; and (iv) 3 improvement compared to REP-CNN [5].…”

“…(M1) the JPEG 2000 codec [29] based on the OpenJPEG implementation [33], the active reference software for JPEG 2000 [34], where the code runs with the "−r 1" parameter for a lossless compression setting; (M2) the HEVC video codec [13] with all intra configuration; HEVC encodes the pseudo-video-sequence created using the spiral stacking scan pattern [10,12] [6], our preliminary work on MP-wise prediction, where the models are trained based on patches selected from the same training set; (M8) the PredNN method [4], the first paper on pixel-wise CNN-based prediction, where the model is trained based on patches collected from the same training set; for each LF image more than 183 million patches are processed, one for each pixel and for each color matrix; (M9) the REP-CNN method [5], the first deep-learningbased dual prediction method for pixel-wise prediction based on a similar training process as M8; (M10) the proposed MPSC-CNN codec employed for the 2 × 2 configuration of reference views; (M11) the proposed MPSC-CNN codec employed for the 3 × 3 configuration of reference views;…”

“…In this paper, the selected reference views are encoded lossless in the first stage as a base layer. The corresponding reference LF image is encoded by employing the pixel-wise REP-CNN method [5] using network models trained for each color matrix.…”

“…For each view configuration, one MPS-CNN model and one PSMP-CNN model are trained for each color channel, during 32 epochs, and using a batch size of 500 patches; all the trained models are available online [32]. 8 ionut schiopu and adrian munteanu Since the reference LF image is encoded using the REP-CNN method [5], one REP-CNN model was trained for each color channel. Hence, in this paper, a total number of 36 models were trained for the proposed experimental setup.…”

“…, 32, where f d = 0.2 is the decay rate, n s = 5 is the decay step, and η 1 = 5 × 10 −4 is the learning rate at the first epoch. The training procedure is similar to [5,6]. Figure 6 shows the synthesis results for three LF images in the dataset.…”