Jie Liang

Jie Liang, Hui Zeng, Lei Zhang

Single image super-resolution (SISR) with generative adversarial networks (GAN) has recently attracted increasing attention due to its potentials to generate rich details. However, the training of GAN is unstable, and it often introduces many perceptually unpleasant artifacts along with the generated details. In this paper, we demonstrate that it is possible to train a GAN-based SISR model which can stably generate perceptually realistic details while inhibiting visual artifacts. Based on the observation that the local statistics (e.g., residual variance) of artifact areas are often different from the areas of perceptually friendly details, we develop a framework to discriminate between GAN-generated artifacts and realistic details, and consequently generate an artifact map to regularize and stabilize the model training process. Our proposed locally discriminative learning (LDL) method is simple yet effective, which can be easily plugged in off-the-shelf SISR methods and boost their performance. Experiments demonstrate that LDL outperforms the state-of-the-art GAN based SISR methods, achieving not only higher reconstruction accuracy but also superior perceptual quality on both synthetic and real-world datasets. Codes and models are available at https://github.com/csjliang/LDL.

Jie Liang, Hui Zeng, Lei Zhang

Efficient and effective real-world image super-resolution (Real-ISR) is a challenging task due to the unknown complex degradation of real-world images and the limited computation resources in practical applications. Recent research on Real-ISR has achieved significant progress by modeling the image degradation space; however, these methods largely rely on heavy backbone networks and they are inflexible to handle images of different degradation levels. In this paper, we propose an efficient and effective degradation-adaptive super-resolution (DASR) network, whose parameters are adaptively specified by estimating the degradation of each input image. Specifically, a tiny regression network is employed to predict the degradation parameters of the input image, while several convolutional experts with the same topology are jointly optimized to specify the network parameters via a non-linear mixture of experts. The joint optimization of multiple experts and the degradation-adaptive pipeline significantly extend the model capacity to handle degradations of various levels, while the inference remains efficient since only one adaptively specified network is used for super-resolving the input image. Our extensive experiments demonstrate that the proposed DASR is not only much more effective than existing methods on handling real-world images with different degradation levels but also efficient for easy deployment. Codes, models and datasets are available at https://github.com/csjliang/DASR.

Du Chen, Jie Liang, Xindong Zhang et al.

How to generate the ground-truth (GT) image is a critical issue for training realistic image super-resolution (Real-ISR) models. Existing methods mostly take a set of high-resolution (HR) images as GTs and apply various degradations to simulate their low-resolution (LR) counterparts. Though great progress has been achieved, such an LR-HR pair generation scheme has several limitations. First, the perceptual quality of HR images may not be high enough, limiting the quality of Real-ISR outputs. Second, existing schemes do not consider much human perception in GT generation, and the trained models tend to produce over-smoothed results or unpleasant artifacts. With the above considerations, we propose a human guided GT generation scheme. We first elaborately train multiple image enhancement models to improve the perceptual quality of HR images, and enable one LR image having multiple HR counterparts. Human subjects are then involved to annotate the high quality regions among the enhanced HR images as GTs, and label the regions with unpleasant artifacts as negative samples. A human guided GT image dataset with both positive and negative samples is then constructed, and a loss function is proposed to train the Real-ISR models. Experiments show that the Real-ISR models trained on our dataset can produce perceptually more realistic results with less artifacts. Dataset and codes can be found at https://github.com/ChrisDud0257/HGGT

Du Chen, Zhengqiang Zhang, Jie Liang et al.

Generative adversarial networks (GAN) and generative diffusion models (DM) have been widely used in real-world image super-resolution (Real-ISR) to enhance the image perceptual quality. However, these generative models are prone to generating visual artifacts and false image structures, resulting in unnatural Real-ISR results. Based on the fact that natural images exhibit high self-similarities, i.e., a local patch can have many similar patches to it in the whole image, in this work we propose a simple yet effective self-similarity loss (SSL) to improve the performance of generative Real-ISR models, enhancing the hallucination of structural and textural details while reducing the unpleasant visual artifacts. Specifically, we compute a self-similarity graph (SSG) of the ground-truth image, and enforce the SSG of Real-ISR output to be close to it. To reduce the training cost and focus on edge areas, we generate an edge mask from the ground-truth image, and compute the SSG only on the masked pixels. The proposed SSL serves as a general plug-and-play penalty, which could be easily applied to the off-the-shelf Real-ISR models. Our experiments demonstrate that, by coupling with SSL, the performance of many state-of-the-art Real-ISR models, including those GAN and DM based ones, can be largely improved, reproducing more perceptually realistic image details and eliminating many false reconstructions and visual artifacts. Codes and supplementary material can be found at https://github.com/ChrisDud0257/SSL

Haisheng Fu, Feng Liang, Jie Liang et al.

Recently, deep learning-based image compression has made signifcant progresses, and has achieved better ratedistortion (R-D) performance than the latest traditional method, H.266/VVC, in both subjective metric and the more challenging objective metric. However, a major problem is that many leading learned schemes cannot maintain a good trade-off between performance and complexity. In this paper, we propose an effcient and effective image coding framework, which achieves similar R-D performance with lower complexity than the state of the art. First, we develop an improved multi-scale residual block (MSRB) that can expand the receptive feld and is easier to obtain global information. It can further capture and reduce the spatial correlation of the latent representations. Second, a more advanced importance map network is introduced to adaptively allocate bits to different regions of the image. Third, we apply a 2D post-quantization flter (PQF) to reduce the quantization error, motivated by the Sample Adaptive Offset (SAO) flter in video coding. Moreover, We fnd that the complexity of encoder and decoder have different effects on image compression performance. Based on this observation, we design an asymmetric paradigm, in which the encoder employs three stages of MSRBs to improve the learning capacity, whereas the decoder only needs one stage of MSRB to yield satisfactory reconstruction, thereby reducing the decoding complexity without sacrifcing performance. Experimental results show that compared to the state-of-the-art method, the encoding and decoding time of the proposed method are about 17 times faster, and the R-D performance is only reduced by less than 1% on both Kodak and Tecnick datasets, which is still better than H.266/VVC(4:4:4) and other recent learning-based methods. Our source code is publicly available at https://github.com/fengyurenpingsheng.

Tianhe Wu, Kede Ma, Jie Liang et al.

While Multimodal Large Language Models (MLLMs) have experienced significant advancement in visual understanding and reasoning, their potential to serve as powerful, flexible, interpretable, and text-driven models for Image Quality Assessment (IQA) remains largely unexplored. In this paper, we conduct a comprehensive and systematic study of prompting MLLMs for IQA. We first investigate nine prompting systems for MLLMs as the combinations of three standardized testing procedures in psychophysics (i.e., the single-stimulus, double-stimulus, and multiple-stimulus methods) and three popular prompting strategies in natural language processing (i.e., the standard, in-context, and chain-of-thought prompting). We then present a difficult sample selection procedure, taking into account sample diversity and uncertainty, to further challenge MLLMs equipped with the respective optimal prompting systems. We assess three open-source and one closed-source MLLMs on several visual attributes of image quality (e.g., structural and textural distortions, geometric transformations, and color differences) in both full-reference and no-reference scenarios. Experimental results show that only the closed-source GPT-4V provides a reasonable account for human perception of image quality, but is weak at discriminating fine-grained quality variations (e.g., color differences) and at comparing visual quality of multiple images, tasks humans can perform effortlessly.

Lingchen Sun, Jie Liang, Shuaizheng Liu et al.

High perceptual quality and low distortion degree are two important goals in image restoration tasks such as super-resolution (SR). Most of the existing SR methods aim to achieve these goals by minimizing the corresponding yet conflicting losses, such as the $\ell_1$ loss and the adversarial loss. Unfortunately, the commonly used gradient-based optimizers, such as Adam, are hard to balance these objectives due to the opposite gradient decent directions of the contradictory losses. In this paper, we formulate the perception-distortion trade-off in SR as a multi-objective optimization problem and develop a new optimizer by integrating the gradient-free evolutionary algorithm (EA) with gradient-based Adam, where EA and Adam focus on the divergence and convergence of the optimization directions respectively. As a result, a population of optimal models with different perception-distortion preferences is obtained. We then design a fusion network to merge these models into a single stronger one for an effective perception-distortion trade-off. Experiments demonstrate that with the same backbone network, the perception-distortion balanced SR model trained by our method can achieve better perceptual quality than its competitors while attaining better reconstruction fidelity. Codes and models can be found at https://github.com/csslc/EA-Adam}{https://github.com/csslc/EA-Adam.

Jian Jin, Xingxing Zhang, Lili Meng et al.

Recently, various view synthesis distortion estimation models have been studied to better serve for 3-D video coding. However, they can hardly model the relationship quantitatively among different levels of depth changes, texture degeneration, and the view synthesis distortion (VSD), which is crucial for rate-distortion optimization and rate allocation. In this paper, an auto-weighted layer representation based view synthesis distortion estimation model is developed. Firstly, the sub-VSD (S-VSD) is defined according to the level of depth changes and their associated texture degeneration. After that, a set of theoretical derivations demonstrate that the VSD can be approximately decomposed into the S-VSDs multiplied by their associated weights. To obtain the S-VSDs, a layer-based representation of S-VSD is developed, where all the pixels with the same level of depth changes are represented with a layer to enable efficient S-VSD calculation at the layer level. Meanwhile, a nonlinear mapping function is learnt to accurately represent the relationship between the VSD and S-VSDs, automatically providing weights for S-VSDs during the VSD estimation. To learn such function, a dataset of VSD and its associated S-VSDs are built. Experimental results show that the VSD can be accurately estimated with the weights learnt by the nonlinear mapping function once its associated S-VSDs are available. The proposed method outperforms the relevant state-of-the-art methods in both accuracy and efficiency. The dataset and source code of the proposed method will be available at https://github.com/jianjin008/.

Haisheng Fu, Feng Liang, Jianping Lin et al.

Recently deep learning-based image compression methods have achieved significant achievements and gradually outperformed traditional approaches including the latest standard Versatile Video Coding (VVC) in both PSNR and MS-SSIM metrics. Two key components of learned image compression are the entropy model of the latent representations and the encoding/decoding network architectures. Various models have been proposed, such as autoregressive, softmax, logistic mixture, Gaussian mixture, and Laplacian. Existing schemes only use one of these models. However, due to the vast diversity of images, it is not optimal to use one model for all images, even different regions within one image. In this paper, we propose a more flexible discretized Gaussian-Laplacian-Logistic mixture model (GLLMM) for the latent representations, which can adapt to different contents in different images and different regions of one image more accurately and efficiently, given the same complexity. Besides, in the encoding/decoding network design part, we propose a concatenated residual blocks (CRB), where multiple residual blocks are serially connected with additional shortcut connections. The CRB can improve the learning ability of the network, which can further improve the compression performance. Experimental results using the Kodak, Tecnick-100 and Tecnick-40 datasets show that the proposed scheme outperforms all the leading learning-based methods and existing compression standards including VVC intra coding (4:4:4 and 4:2:0) in terms of the PSNR and MS-SSIM. The source code is available at \url{https://github.com/fengyurenpingsheng}

Jie Liang, Hui Zeng, Lei Zhang

Existing image-to-image translation (I2IT) methods are either constrained to low-resolution images or long inference time due to their heavy computational burden on the convolution of high-resolution feature maps. In this paper, we focus on speeding-up the high-resolution photorealistic I2IT tasks based on closed-form Laplacian pyramid decomposition and reconstruction. Specifically, we reveal that the attribute transformations, such as illumination and color manipulation, relate more to the low-frequency component, while the content details can be adaptively refined on high-frequency components. We consequently propose a Laplacian Pyramid Translation Network (LPTN) to simultaneously perform these two tasks, where we design a lightweight network for translating the low-frequency component with reduced resolution and a progressive masking strategy to efficiently refine the high-frequency ones. Our model avoids most of the heavy computation consumed by processing high-resolution feature maps and faithfully preserves the image details. Extensive experimental results on various tasks demonstrate that the proposed method can translate 4K images in real-time using one normal GPU while achieving comparable transformation performance against existing methods. Datasets and codes are available: https://github.com/csjliang/LPTN.

Jinbo Yan, Rui Peng, Zhiyan Wang et al.

Building Free-Viewpoint Videos in a streaming manner offers the advantage of rapid responsiveness compared to offline training methods, greatly enhancing user experience. However, current streaming approaches face challenges of high per-frame reconstruction time (10s+) and error accumulation, limiting their broader application. In this paper, we propose Instant Gaussian Stream (IGS), a fast and generalizable streaming framework, to address these issues. First, we introduce a generalized Anchor-driven Gaussian Motion Network, which projects multi-view 2D motion features into 3D space, using anchor points to drive the motion of all Gaussians. This generalized Network generates the motion of Gaussians for each target frame in the time required for a single inference. Second, we propose a Key-frame-guided Streaming Strategy that refines each key frame, enabling accurate reconstruction of temporally complex scenes while mitigating error accumulation. We conducted extensive in-domain and cross-domain evaluations, demonstrating that our approach can achieve streaming with a average per-frame reconstruction time of 2s+, alongside a enhancement in view synthesis quality.

Haisheng Fu, Jie Liang, Feng Liang et al.

Learned image compression (LIC) has recently made significant progress, surpassing traditional methods. However, most LIC approaches operate mainly in the spatial domain and lack mechanisms for reducing frequency-domain correlations. To address this, we propose a novel framework that integrates low-complexity 3D multi-level Discrete Wavelet Transform (DWT) into convolutional layers and entropy coding, reducing both spatial and channel correlations to improve frequency selectivity and rate-distortion (R-D) performance. Our proposed 3D multi-level wavelet-domain convolution (3DM-WeConv) layer first applies 3D multi-level DWT (e.g., 5/3 and 9/7 wavelets from JPEG 2000) to transform data into the wavelet domain. Then, different-sized convolutions are applied to different frequency subbands, followed by inverse 3D DWT to restore the spatial domain. The 3DM-WeConv layer can be flexibly used within existing CNN-based LIC models. We also introduce a 3D wavelet-domain channel-wise autoregressive entropy model (3DWeChARM), which performs slice-based entropy coding in the 3D DWT domain. Low-frequency (LF) slices are encoded first to provide priors for high-frequency (HF) slices. A two-step training strategy is adopted: first balancing LF and HF rates, then fine-tuning with separate weights. Extensive experiments demonstrate that our framework consistently outperforms state-of-the-art CNN-based LIC methods in R-D performance and computational complexity, with larger gains for high-resolution images. On the Kodak, Tecnick 100, and CLIC test sets, our method achieves BD-Rate reductions of -12.24%, -15.51%, and -12.97%, respectively, compared to H.266/VVC.

Boshen Wang, Bowei Ye, Lin Xu et al.

In silico methods evaluating the mutation effects of missense mutations are providing an important approach for understanding mutations in personal genomes and identifying disease-relevant biomarkers. However, existing methods, including deep learning methods, heavily rely on sequence-aware information, and do not fully leverage the potential of available 3D structural information. In addition, these methods may exhibit an inability to predict mutations in domains difficult to formulate sequence-based embeddings. In this study, we introduce a novel rationale-guided graph neural network AlphaGMut to evaluate mutation effects and to distinguish pathogenic mutations from neutral mutations. We compute the alpha shapes of protein structures to obtain atomic-resolution edge connectivities and map them to an accurate residue-level graph representation. We then compute structural-, topological-, biophysical-, and sequence properties of the mutation sites, which are assigned as node attributes in the graph. These node attributes could effectively guide the graph neural network to learn the difference between pathogenic and neutral mutations using k-hop message passing with a short training period. We demonstrate that AlphaGMut outperforms state-of-the-art methods, including DeepMind's AlphaMissense, in many performance metrics. In addition, AlphaGMut has the advantage of performing well in alignment-free settings, which provides broader prediction coverage and better generalization compared to current methods requiring deep sequence-aware information.

Xinyi Liu, Qian Zhao, Jie Liang et al.

Guided image restoration (GIR), such as guided depth map super-resolution and pan-sharpening, aims to enhance a target image using guidance information from another image of the same scene. Currently, joint image filtering-inspired deep learning-based methods represent the state-of-the-art for GIR tasks. Those methods either deal with GIR in an end-to-end way by elaborately designing filtering-oriented deep neural network (DNN) modules, focusing on the feature-level fusion of inputs; or explicitly making use of the traditional joint filtering mechanism by parameterizing filtering coefficients with DNNs, working on image-level fusion. The former ones are good at recovering contextual information but tend to lose fine-grained details, while the latter ones can better retain textual information but might lead to content distortions. In this work, to inherit the advantages of both methodologies while mitigating their limitations, we proposed a Simultaneous Feature and Image Guided Fusion (SFIGF) network, that simultaneously considers feature and image-level guided fusion following the guided filter (GF) mechanism. In the feature domain, we connect the cross-attention (CA) with GF, and propose a GF-inspired CA module for better feature-level fusion; in the image domain, we fully explore the GF mechanism and design GF-like structure for better image-level fusion. Since guided fusion is implemented in both feature and image domains, the proposed SFIGF is expected to faithfully reconstruct both contextual and textual information from sources and thus lead to better GIR results. We apply SFIGF to 4 typical GIR tasks, and experimental results on these tasks demonstrate its effectiveness and general availability.

Binglin Li, Jie Liang, Haisheng Fu et al.

Encoding the Region Of Interest (ROI) with better quality than the background has many applications including video conferencing systems, video surveillance and object-oriented vision tasks. In this paper, we propose a ROI-based image compression framework with Swin transformers as main building blocks for the autoencoder network. The binary ROI mask is integrated into different layers of the network to provide spatial information guidance. Based on the ROI mask, we can control the relative importance of the ROI and non-ROI by modifying the corresponding Lagrange multiplier $ λ$ for different regions. Experimental results show our model achieves higher ROI PSNR than other methods and modest average PSNR for human evaluation. When tested on models pre-trained with original images, it has superior object detection and instance segmentation performance on the COCO validation dataset.

Mohammad Akbari, Jie Liang, Jingning Han et al.

Recently deep learning-based image compression has shown the potential to outperform traditional codecs. However, most existing methods train multiple networks for multiple bit rates, which increase the implementation complexity. In this paper, we propose a new variable-rate image compression framework, which employs generalized octave convolutions (GoConv) and generalized octave transposed-convolutions (GoTConv) with built-in generalized divisive normalization (GDN) and inverse GDN (IGDN) layers. Novel GoConv- and GoTConv-based residual blocks are also developed in the encoder and decoder networks. Our scheme also uses a stochastic rounding-based scalar quantization. To further improve the performance, we encode the residual between the input and the reconstructed image from the decoder network as an enhancement layer. To enable a single model to operate with different bit rates and to learn multi-rate image features, a new objective function is introduced. Experimental results show that the proposed framework trained with variable-rate objective function outperforms the standard codecs such as H.265/HEVC-based BPG and state-of-the-art learning-based variable-rate methods.

Ye Li, Kangning Yin, Jie Liang et al.

Person search generally involves three important parts: person detection, feature extraction and identity comparison. However, person search integrating detection, extraction and comparison has the following drawbacks. Firstly, the accuracy of detection will affect the accuracy of comparison. Secondly, it is difficult to achieve real-time in real-world applications. To solve these problems, we propose a Multi-task Joint Framework for real-time person search (MJF), which optimizes the person detection, feature extraction and identity comparison respectively. For the person detection module, we proposed the YOLOv5-GS model, which is trained with person dataset. It combines the advantages of the Ghostnet and the Squeeze-and-Excitation (SE) block, and improves the speed and accuracy. For the feature extraction module, we design the Model Adaptation Architecture (MAA), which could select different network according to the number of people. It could balance the relationship between accuracy and speed. For identity comparison, we propose a Three Dimension (3D) Pooled Table and a matching strategy to improve identification accuracy. On the condition of 1920*1080 resolution video and 500 IDs table, the identification rate (IR) and frames per second (FPS) achieved by our method could reach 93.6% and 25.7,

Mohammad Akbari, Jie Liang, Jingning Han et al.

Learned image compression has recently shown the potential to outperform the standard codecs. State-of-the-art rate-distortion (R-D) performance has been achieved by context-adaptive entropy coding approaches in which hyperprior and autoregressive models are jointly utilized to effectively capture the spatial dependencies in the latent representations. However, the latents are feature maps of the same spatial resolution in previous works, which contain some redundancies that affect the R-D performance. In this paper, we propose the first learned multi-frequency image compression and entropy coding approach that is based on the recently developed octave convolutions to factorize the latents into high and low frequency (resolution) components, where the low frequency is represented by a lower resolution. Therefore, its spatial redundancy is reduced, which improves the R-D performance. Novel generalized octave convolution and octave transposed-convolution architectures with internal activation layers are also proposed to preserve more spatial structure of the information. Experimental results show that the proposed scheme not only outperforms all existing learned methods as well as standard codecs such as the next-generation video coding standard VVC (4:2:0) on the Kodak dataset in both PSNR and MS-SSIM. We also show that the proposed generalized octave convolution can improve the performance of other auto-encoder-based computer vision tasks such as semantic segmentation and image denoising.

Binglin Li, Mohammad Akbari, Jie Liang et al.

Recently many works attempt to develop image compression models based on deep learning architectures, where the uniform scalar quantizer (SQ) is commonly applied to the feature maps between the encoder and decoder. In this paper, we propose to incorporate trellis coded quantizer (TCQ) into a deep learning based image compression framework. A soft-to-hard strategy is applied to allow for back propagation during training. We develop a simple image compression model that consists of three subnetworks (encoder, decoder and entropy estimation), and optimize all of the components in an end-to-end manner. We experiment on two high resolution image datasets and both show that our model can achieve superior performance at low bit rates. We also show the comparisons between TCQ and SQ based on our proposed baseline model and demonstrate the advantage of TCQ.

Mohammad Akbari, Jie Liang, Jingning Han et al.

Recently it has been shown that deep learning-based image compression has shown the potential to outperform traditional codecs. However, most existing methods train multiple networks for multiple bit rates, which increases the implementation complexity. In this paper, we propose a variable-rate image compression framework, which employs more Generalized Divisive Normalization (GDN) layers than previous GDN-based methods. Novel GDN-based residual sub-networks are also developed in the encoder and decoder networks. Our scheme also uses a stochastic rounding-based scalable quantization. To further improve the performance, we encode the residual between the input and the reconstructed image from the decoder network as an enhancement layer. To enable a single model to operate with different bit rates and to learn multi-rate image features, a new objective function is introduced. Experimental results show that the proposed framework trained with variable-rate objective function outperforms all standard codecs such as H.265/HEVC-based BPG and state-of-the-art learning-based variable-rate methods.

Haisheng Fu, Feng Liang, Bo Lei et al.

Recently deep learning-based methods have been applied in image compression and achieved many promising results. In this paper, we propose an improved hybrid layered image compression framework by combining deep learning and the traditional image codecs. At the encoder, we first use a convolutional neural network (CNN) to obtain a compact representation of the input image, which is losslessly encoded by the FLIF codec as the base layer of the bit stream. A coarse reconstruction of the input is obtained by another CNN from the reconstructed compact representation. The residual between the input and the coarse reconstruction is then obtained and encoded by the H.265/HEVC-based BPG codec as the enhancement layer of the bit stream. Experimental results using the Kodak and Tecnick datasets show that the proposed scheme outperforms the state-of-the-art deep learning-based layered coding scheme and traditional codecs including BPG in both PSNR and MS-SSIM metrics across a wide range of bit rates, when the images are coded in the RGB444 domain.

Mohammad Akbari, Jie Liang, Jingning Han

Deep learning has revolutionized many computer vision fields in the last few years, including learning-based image compression. In this paper, we propose a deep semantic segmentation-based layered image compression (DSSLIC) framework in which the semantic segmentation map of the input image is obtained and encoded as the base layer of the bit-stream. A compact representation of the input image is also generated and encoded as the first enhancement layer. The segmentation map and the compact version of the image are then employed to obtain a coarse reconstruction of the image. The residual between the input and the coarse reconstruction is additionally encoded as another enhancement layer. Experimental results show that the proposed framework outperforms the H.265/HEVC-based BPG and other codecs in both PSNR and MS-SSIM metrics across a wide range of bit rates in RGB domain. Besides, since semantic segmentation map is included in the bit-stream, the proposed scheme can facilitate many other tasks such as image search and object-based adaptive image compression.

Mohammad Akbari, Jie Liang

A semi-recurrent hybrid VAE-GAN model for generating sequential data is introduced. In order to consider the spatial correlation of the data in each frame of the generated sequence, CNNs are utilized in the encoder, generator, and discriminator. The subsequent frames are sampled from the latent distributions obtained by encoding the previous frames. As a result, the dependencies between the frames are maintained. Two testing frameworks for synthesizing a sequence with any number of frames are also proposed. The promising experimental results on piano music generation indicates the potential of the proposed framework in modeling other sequential data such as video.

Xing Wang, Jie Liang

Deep neural networks generally involve some layers with mil- lions of parameters, making them difficult to be deployed and updated on devices with limited resources such as mobile phones and other smart embedded systems. In this paper, we propose a scalable representation of the network parameters, so that different applications can select the most suitable bit rate of the network based on their own storage constraints. Moreover, when a device needs to upgrade to a high-rate network, the existing low-rate network can be reused, and only some incremental data are needed to be downloaded. We first hierarchically quantize the weights of a pre-trained deep neural network to enforce weight sharing. Next, we adaptively select the bits assigned to each layer given the total bit budget. After that, we retrain the network to fine-tune the quantized centroids. Experimental results show that our method can achieve scalable compression with graceful degradation in the performance.