Toward Intelligent Sensing: Intermediate Deep Feature Compression

Chen, Zhuo; Fan, Kui; Wang, Shiqi; Duan, Ling-Yu; Wang, Lin; Kot, Alex C.

doi:10.1109/tip.2019.2941660

Cited by 105 publications

(52 citation statements)

References 43 publications

(81 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Authors in [6,24] describe general lossy and lossless codecs for deep feature compression. Their focus is on the features of popular DNN backbones rather than task-specific features.…”

Section: Related Workmentioning

confidence: 99%

“…Due to their focus on single tensor compression, none of the studies mentioned above consider optimal bit allocation to multiple tensors. Even in [6,24] where compression of multiple features is considered, the compression is performed without joint bit allocation. The main contribution of the present paper are the solutions to bit allocation problems in several multi-stream CI scenarios.…”

Section: Related Workmentioning

confidence: 99%

“…To efficiently utilize the communication channel in CI, the feature tensors obtained from the edge sub-model need to be compressed prior to transmission to the cloud [6]. For singlestream DNNs such as VGG [7], only one feature tensor needs to be compressed, so the entire available bit budget is assigned to this tensor.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Bit Allocation for Multi-Task Collaborative Intelligence

Alvar

Bajić

2020

ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

Recent studies have shown that collaborative intelligence (CI) is a promising framework for deployment of Artificial Intelligence (AI)-based services on mobile devices. In CI, a deep neural network is split between the mobile device and the cloud. Deep features obtained at the mobile are compressed and transferred to the cloud to complete the inference. So far, the methods in the literature focused on transferring a single deep feature tensor from the mobile to the cloud. Such methods are not applicable to some recent, high-performance networks with multiple branches and skip connections. In this paper, we propose the first bit allocation method for multistream, multi-task CI. We first establish a model for the joint distortion of the multiple tasks as a function of the bit rates assigned to different deep feature tensors. Then, using the proposed model, we solve the rate-distortion optimization problem under a total rate constraint to obtain the best rate allocation among the tensors to be transferred. Experimental results illustrate the efficacy of the proposed scheme compared to several alternative bit allocation methods.

show abstract

“…Authors in [6,24] describe general lossy and lossless codecs for deep feature compression. Their focus is on the features of popular DNN backbones rather than task-specific features.…”

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Bit Allocation for Multi-Task Collaborative Intelligence

Alvar

Bajić

2020

ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

show abstract

“…Currently, there are several directions of lossless compression algorithm development, including lossless algorithm improvement, data encryption, and lossless deep learning feature compression [21]- [23]. Compression algorithm such as the arithmetic coding method can be improved by using the range adjusting, step size adjusting, and mutual learning scheme [21], where the compression efficiency is mainly focused.…”

Section: Introductionmentioning

confidence: 99%

“…In addition, due to recent advances of hardware technology, the intelligent analysis equipped at the front-end with deep learning becomes practical. Report [23] proposes a strategy to compactly represent and convey the intermediate-layer deep learning features with high generalization capability, to facilitate the collaborating approach between front and cloud ends. Thus, lossless compression of deep learning features demonstrates a promising feasibility, as a series of tasks can simultaneously benefit from the transmitted intermediate layer features.…”

Section: Introductionmentioning

confidence: 99%

2-D Bi-Level Block Coding for Color Image Compression and Transmission With Bit-Error Awareness

et al. 2020

View full text Add to dashboard Cite

In this paper, we develop a new color image lossless compression algorithm with bit-error awareness based on a general bi-level block coding method. The proposed method contains three stages. First, a color image in the RGB color space is converted to the YCrCb color space by lossless reversible transformations. Next, predictors in the YCrCb color space are applied to generate the residue sequences. In order to achieve a better image compression ratio, a particle swarm optimization (PSO) algorithm is adopted to search the best combination from the candidates of color transformations and predictors which generate the minimum residue entropy. Third, a new 2-D bi-level block coding is developed to further encode the residue sequences. Comparing with the existing residue coding methods, including 1-D bi-level block coding, interval Huffman coding, and standard Huffman coding, the 2-D bi-level block coding algorithm can improve image compression ratio as well as preserving the bit-error resilience. Finally, the key parameters such as color transformation information, predictor parameters and residue coding parameters are protected using (7, 4) Hamming code in the bit stream before transmission. The performances are validated in terms of compression ratio (CR) and peak signal to noise ratio (PSNR). The compression algorithm with bi-level coding achieves the highest PSNR when the bit-error rate (BER) is larger than 0.001 and maintains an acceptable PSNR for BER less than 0.001. In particular, the compression algorithm using the new 2-D bi-level block coding scheme achieves the highest CR. INDEX TERMS lossless compression, bit-error awareness, general bi-level block coding, particle swarm optimization, (7, 4) Hamming code.

show abstract